Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
477 lines (433 loc) · 45.7 KB
/
Copy pathindex.html
File metadata and controls
477 lines (433 loc) · 45.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Rajatendu Dey - Backend and Identity Engineer</title>
<meta name="description" content="Backend and identity engineer, formerly at AWS and HSBC, with an earlier machine learning internship at HighRadius. A growing series of tested, measured projects, starting with a live LLM gateway.">
<meta property="og:title" content="Rajatendu Dey - Backend and Identity Engineer">
<meta property="og:description" content="Tested, measured projects with live demos: an LLM gateway that keeps apps answering when providers fail, and delegation tokens that narrow what AI agents may do.">
<meta property="og:type" content="website">
<meta property="og:url" content="https://rajatendu-dey.vercel.app/">
<meta property="og:image" content="https://rajatendu-dey.vercel.app/og.png">
<link rel="icon" href="favicon.svg" type="image/svg+xml">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;600&display=swap" rel="stylesheet">
<link rel="stylesheet" href="style.css">
</head>
<body>
<nav class="nav">
<div class="nav-inner">
<a class="nav-name" href="#top">Rajatendu Dey</a>
<div class="nav-links">
<a href="#projects">Projects</a>
<a href="#about">About</a>
<a href="#contact">Contact</a>
</div>
</div>
</nav>
<main id="top">
<header class="hero reveal">
<div class="hero-grid">
<div class="hero-id">
<p class="hero-eyebrow mono"><span class="live-dot"></span>Backend and identity engineer</p>
<h1>Rajatendu Dey</h1>
<p class="hero-summary">I build <strong>distributed backend systems, authentication and identity infrastructure, and AI-powered applications</strong>. Most recently SDE II on the AWS UnifiedAuth platform; previously an engineer at HSBC, with an earlier machine learning internship at HighRadius.</p>
<p class="hero-meta">Hooghly, India · Open to Bangalore and remote roles</p>
<div class="hero-links">
<a class="button button-primary" href="mailto:rajatendud@gmail.com">Email me</a>
<a class="button" href="https://linkedin.com/in/rajatendu-dey" target="_blank" rel="noopener">LinkedIn</a>
<a class="button" href="https://github.com/Rajatendu1" target="_blank" rel="noopener">GitHub</a>
<a class="button" href="#open-source">Open source</a>
<a class="button" href="https://github.com/sponsors/Rajatendu1" target="_blank" rel="noopener">Sponsor</a>
</div>
</div>
<aside class="hero-panel" aria-label="Live projects">
<p class="hero-panel-label mono"><span class="live-dot"></span>Live right now</p>
<div class="hero-project">
<div class="hero-project-head">
<span class="hero-project-name">llmgate</span>
<span class="hero-project-tag mono">LLM infrastructure</span>
</div>
<p class="hero-project-line">One front door for AI models, with cautious automatic failover when a provider stumbles.</p>
<a class="hero-try mono" href="https://llmgate-demo.vercel.app" target="_blank" rel="noopener">Try the live demo<span class="hero-try-arrow">-></span></a>
</div>
<div class="hero-project">
<div class="hero-project-head">
<span class="hero-project-name">actas</span>
<span class="hero-project-tag mono">Identity</span>
</div>
<p class="hero-project-line">Tokens for AI agents that can only lose permissions, never gain them.</p>
<a class="hero-try mono" href="https://actas-rajatendu.vercel.app" target="_blank" rel="noopener">Try the live demo<span class="hero-try-arrow">-></span></a>
</div>
<p class="hero-panel-note">Demos run on free tiers. Simulated by default; real-model buttons are marked.</p>
</aside>
</div>
</header>
<section id="projects" class="section reveal">
<p class="section-label mono"><span class="section-number">01</span> Projects</p>
<p class="section-intro">Every entry ships with an evidence summary: tests, measured results and a live demo where one exists. Public source: <a class="text-link" href="https://github.com/Rajatendu1/llmgate">llmgate</a>, <a class="text-link" href="https://github.com/Rajatendu1/actas">Actas</a> and <a class="text-link" href="https://github.com/Rajatendu1/paper-trading-india">paper-trading-india</a>.</p>
<article class="project-card reveal">
<div class="project-head">
<div>
<p class="project-eyebrow mono">Project 01 · LLM infrastructure</p>
<h2>llmgate</h2>
</div>
<a class="pill mono" href="https://llmgate-demo.vercel.app" target="_blank" rel="noopener"><span class="pill-dot"></span>Live demo</a>
</div>
<p class="project-lede">An OpenAI-compatible API gateway in Python and FastAPI: routing, failover, streaming, per-key rate limits and usage metering, deployed live.</p>
<p class="hook">One reliable front door for AI models: when a provider slows down or fails safely, the app behind it keeps answering on a backup. Ambiguous failures are surfaced, never silently retried.</p>
<figure class="diagram">
<svg viewBox="0 0 480 560" role="img" aria-label="How a request flows through llmgate: key auth and rate limit, then routing with circuit breakers, to the primary provider or the backup on failover; usage is metered after the response">
<defs>
<marker id="arr" viewBox="0 0 8 8" refX="7" refY="4" markerWidth="7" markerHeight="7" orient="auto-start-reverse"><path d="M0,0 L8,4 L0,8 z" fill="#94a0b0"/></marker>
<marker id="arrA" viewBox="0 0 8 8" refX="7" refY="4" markerWidth="7" markerHeight="7" orient="auto-start-reverse"><path d="M0,0 L8,4 L0,8 z" fill="#5eead4"/></marker>
</defs>
<rect x="170" y="8" width="140" height="42" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="34" text-anchor="middle" font-family="Inter,sans-serif" font-size="15" font-weight="600" fill="#e8ecf1">Client</text>
<line class="flow " x1="240" y1="50" x2="240" y2="92" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr)"/>
<text x="254" y="78" font-family="'JetBrains Mono',monospace" font-size="11" fill="#94a0b0">POST /v1/chat/completions</text>
<rect x="60" y="96" width="360" height="196" rx="14" fill="rgba(94,234,212,0.07)" stroke="rgba(94,234,212,0.3)"/>
<text x="78" y="120" font-family="'JetBrains Mono',monospace" font-size="12" font-weight="600" fill="#5eead4">llmgate · admission</text>
<rect x="90" y="130" width="300" height="58" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="153" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Key auth + rate limit</text>
<text x="240" y="171" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">Redis · atomic Lua take · 429 + Retry-After</text>
<line class="flow " x1="240" y1="188" x2="240" y2="202" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr)"/>
<rect x="90" y="206" width="300" height="72" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="229" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Router</text>
<text x="240" y="247" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">priority tiers · weighted routing</text>
<text x="240" y="263" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">circuit breaker per provider</text>
<line class="flow " x1="170" y1="292" x2="140" y2="336" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr)"/>
<line x1="310" y1="292" x2="340" y2="336" stroke="#5eead4" stroke-width="1.5" class="flow-slow" marker-end="url(#arrA)"/>
<text x="118" y="322" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">healthy</text>
<text x="352" y="322" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#5eead4">failover</text>
<rect x="40" y="340" width="190" height="56" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="135" y="364" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Provider A</text>
<text x="135" y="382" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">primary</text>
<rect x="250" y="340" width="190" height="56" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="345" y="364" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Provider B</text>
<text x="345" y="382" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">backup · simulated</text>
<line class="flow " x1="240" y1="396" x2="240" y2="434" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr)"/>
<text x="254" y="420" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">response</text>
<rect x="90" y="438" width="300" height="70" rx="10" fill="rgba(94,234,212,0.07)" stroke="rgba(94,234,212,0.3)"/>
<text x="240" y="460" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Usage metering</text>
<text x="240" y="477" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">usage settles after the response</text>
<text x="240" y="493" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">requests · tokens · est. cost per key</text>
<text x="240" y="540" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">real Groq primary + backup in the demo; failover exercise stays simulated</text>
</svg>
<div class="diagram-steps" aria-hidden="true">
<div class="dstep"><p class="dstep-title">Client</p><p class="dstep-sub mono">POST /v1/chat/completions</p></div>
<div class="darrow">↓</div>
<div class="dstep tint"><p class="dstep-tag mono">llmgate · admission</p><p class="dstep-title">Key auth + rate limit</p><p class="dstep-sub mono">Redis · atomic Lua take · 429 + Retry-After</p></div>
<div class="darrow">↓</div>
<div class="dstep tint"><p class="dstep-title">Router</p><p class="dstep-sub mono">priority tiers · weighted routing · circuit breaker per provider</p></div>
<div class="darrow">↓ healthy</div>
<div class="dstep"><p class="dstep-title">Provider A · primary</p></div>
<div class="darrow">↓ on failure: failover</div>
<div class="dstep"><p class="dstep-title">Provider B · backup (simulated)</p></div>
<div class="darrow">↓ response</div>
<div class="dstep tint"><p class="dstep-tag mono">llmgate · settlement</p><p class="dstep-title">Usage metering</p><p class="dstep-sub mono">requests · tokens · est. cost per key</p></div>
<p class="dnote mono">real Groq primary + backup in the demo · failover exercise stays simulated</p>
</div>
<figcaption>The request path as implemented. Bounded retries only for classified retryable failures; the live demo exercises availability-first failover against simulated upstreams, while explicit buttons call real Groq models alongside the default simulated Try.</figcaption>
</figure>
<div class="try-row">
<a class="button button-primary" href="https://llmgate-demo.vercel.app" target="_blank" rel="noopener">Try the live demo</a>
<p class="try-note">One click sends a request through the gateway and shows the route it took; another hands work to a sub-agent and shows it being refused a model outside its permissions. Or mint a 24-hour key and call the gateway from your own OpenAI-compatible client - Python and Node quickstarts included; minted keys reach the simulated fleet only.</p>
</div>
<p class="detail-label mono">The engineering</p>
<ul class="feature-list">
<li>One <code>/v1/chat/completions</code> endpoint in front of several providers, with priority tiers, weighted routing and a circuit breaker per provider.</li>
<li>Retries classified by generation risk: bounded same-provider retries with backoff only for classified retryable failures (connect-phase failure, or a 429 treated as rejected at admission, a documented policy assumption rather than a provider guarantee).</li>
<li>Ambiguous failures (timeouts, 5xx, mid-stream) are never retried in place; failover after them is duplicate-averse by default. The live demo exercises availability-first failover against simulated upstreams; explicit real-model buttons call live Groq models (a primary and a direct-call backup) next to the default simulated Try, with real failover verified locally against fault injection, not an actual outage.</li>
<li>Delegation built in: an embedded Actas token exchange (RFC 8693) narrows a user token down a chain of agents - the sub-agent reaches the plain simulated model but is refused the expensive one with 403 insufficient_scope, a scope-widening attempt is rejected at the exchange, and every grant and denial lands in a hash-chained audit log whose integrity is verified before the result renders.</li>
<li>Measured stage timings: every response carries a standard Server-Timing header with measured spans - authentication, admission (limiter, scope, token bucket), routing (candidate order, breaker, retry backoff), the provider call (time to first chunk for streams) and usage metering. The demo shows these real per-request measurements against each stage of the request journey; the flow animation stays illustrative and is labelled as such. Per request at 0.1 ms granularity, a view of one request rather than a benchmark: streamed responses carry pre-body spans only, and a refusal stops at the refusing stage.</li>
<li>Per-key request and token rate limits on shared Redis, one atomic Lua call per take, with 429 and Retry-After; fail-closed if the store is down.</li>
<li>API keys hashed at rest and rotated live; Prometheus metrics, request IDs, security headers and a dependency-aware readiness probe.</li>
</ul>
<div class="metrics">
<div class="metric">
<p class="metric-value mono">276 <span>rps</span></p>
<p class="metric-label">Healthy, non-stream</p>
<p class="metric-sub mono">p99 364 ms · 4,000 req · all 200</p>
</div>
<div class="metric">
<p class="metric-value mono">248 <span>rps</span></p>
<p class="metric-label">Healthy, streaming</p>
<p class="metric-sub mono">p99 395 ms · 2,000 req · all 200</p>
</div>
<div class="metric">
<p class="metric-value mono">575 <span>rps</span></p>
<p class="metric-label">Primary fails 30%</p>
<p class="metric-sub mono">all 200 · backup served 1,265</p>
</div>
<div class="metric">
<p class="metric-value mono">588 <span>rps</span></p>
<p class="metric-label">Primary fully down</p>
<p class="metric-sub mono">all 200 · every call via backup</p>
</div>
</div>
<p class="metrics-note">Historical local load-test results recorded 9 Nov 2026 against simulated providers, 32 clients. Not rerun against the current build; a sanity check, not a capacity claim.</p>
<dl class="facts">
<div><dt>Verification</dt><dd>Historical result recorded 9 Nov 2026: 162 unit, API and integration tests, 61 end-to-end shell checks and a 23-check JS smoke gate; real-model and delegation demo checks were recorded that day. Current checks and results are in the public repository.</dd></div>
<div><dt>Stack</dt><dd>Python, FastAPI, Redis (Upstash), Vercel serverless, Docker packaging</dd></div>
</dl>
<div class="limits">
<p class="limits-title">Honest limits</p>
<p>The deployment can configure real Groq routes; configuration does not establish current provider availability or pricing. Historical demo measurements recorded 9 Nov 2026 follow, not measurements of the current build. The default Try button is simulated; the real-model button on the primary is verified at 78 prompt + 16 completion tokens in about 415 ms; a direct backup button (a direct call, not failover) returned 78 + 19 tokens in about 180 ms, and both real buttons share one global budget. Real failover is verified only locally: a fault-injected dead primary was retried once, then the real backup answered (78 + 14 tokens, 504 ms) - not an actual provider outage, and independent-vendor failover remains open. The benchmarks and the failover load exercise remain simulated. The real buttons are rate-limited (about one request per 100 seconds, small bursts; a 429 means it is recharging). The delegation demo embeds the issuer inside the gateway process (one trust domain: no remote issuer or JWKS) and uses explicitly unbound bearer tokens - proof-bound (cnf) tokens are rejected rather than downgraded, and each run issues and consumes its own tokens; model allowlists and budgets come from server configuration, never from token claims. It demonstrates delegation semantics and least-privilege enforcement, not production authorization. Breaker and latency state is per instance. Runs on Vercel Hobby (personal, non-commercial); the Redis plan is unverified. Usage totals are operational telemetry, not billing-grade. The Docker image is unbuilt. A prototype, not production-certified.</p>
</div>
<div class="project-links">
<a class="text-link" href="https://llmgate-demo.vercel.app" target="_blank" rel="noopener">Open the live demo</a>
<p class="project-links-note">Try it live: the default button sends a fixed prompt through the full gateway path to simulated providers; explicit buttons call the real Groq primary and the real backup directly (a direct call, not failover); and a delegation button runs a complete RFC 8693 exchange - a sub-agent reaches the cheap model, is refused the expensive one with 403, and a widening attempt is rejected. The real buttons share a tight budget (about one request per 100 seconds; a 429 means it is recharging).</p>
</div>
</article>
<article class="project-card reveal">
<div class="project-head">
<div>
<p class="project-eyebrow mono">Project 02 · Agent identity</p>
<h2>actas</h2>
</div>
<a class="pill mono" href="https://actas-rajatendu.vercel.app" target="_blank" rel="noopener"><span class="pill-dot"></span>Live demo</a>
</div>
<p class="project-lede">Delegation tokens for AI agents: a reference implementation of RFC 8693 (OAuth 2.0 Token Exchange), built from public specs.</p>
<p class="hook">When one AI agent hands work to another, actas makes sure the permissions can only shrink, never grow - and a stolen token is useless without the key it was bound to.</p>
<figure class="diagram">
<svg viewBox="0 0 480 570" role="img" aria-label="Actas delegation story: broad user grant narrows at each hop, a widening attempt is denied, a stolen token without its key is rejected, and every outcome reaches the audit log">
<defs>
<marker id="arr2" viewBox="0 0 8 8" refX="7" refY="4" markerWidth="7" markerHeight="7" orient="auto-start-reverse"><path d="M0,0 L8,4 L0,8 z" fill="#94a0b0"/></marker>
</defs>
<rect x="170" y="8" width="140" height="42" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="34" text-anchor="middle" font-family="Inter,sans-serif" font-size="15" font-weight="600" fill="#e8ecf1">Asha · user</text>
<line class="flow " x1="240" y1="50" x2="240" y2="92" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr2)"/>
<text x="252" y="64" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">read:calendar read:profile</text>
<text x="252" y="80" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">write:notes · 3600s</text>
<rect x="150" y="96" width="180" height="50" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="118" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">planner-bot</text>
<text x="240" y="134" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">aud: calendar-api · 600s</text>
<line class="flow " x1="240" y1="146" x2="240" y2="188" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr2)"/>
<text x="252" y="162" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">keeps: read:calendar write:notes</text>
<text x="252" y="178" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0" text-decoration="line-through">drops: read:profile</text>
<rect x="150" y="192" width="180" height="50" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="214" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">summarizer-tool</text>
<text x="240" y="230" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">nested delegation · 300s</text>
<line class="flow " x1="240" y1="242" x2="240" y2="284" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr2)"/>
<text x="252" y="258" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">keeps: read:calendar</text>
<text x="252" y="274" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0" text-decoration="line-through">drops: write:notes</text>
<rect x="140" y="288" width="200" height="50" rx="10" fill="rgba(255,255,255,0.05)" stroke="#f87171" stroke-opacity="0.7"/>
<text x="240" y="310" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">widening attempt</text>
<text x="240" y="326" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10" fill="#f87171">invalid_scope: admin:all denied</text>
<line class="flow " x1="240" y1="338" x2="240" y2="372" stroke="#94a0b0" stroke-width="1.5" marker-end="url(#arr2)"/>
<rect x="140" y="376" width="200" height="50" rx="10" fill="rgba(255,255,255,0.05)" stroke="rgba(255,255,255,0.12)"/>
<text x="240" y="398" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">key-bound token (cnf.jkt)</text>
<text x="240" y="414" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10" fill="#94a0b0">stolen token, no key: rejected</text>
<line x1="240" y1="426" x2="240" y2="448" stroke="#94a0b0" stroke-width="1" stroke-dasharray="2 4"/>
<rect x="60" y="452" width="360" height="64" rx="10" fill="rgba(94,234,212,0.07)" stroke="rgba(94,234,212,0.3)"/>
<text x="240" y="476" text-anchor="middle" font-family="Inter,sans-serif" font-size="14" font-weight="600" fill="#e8ecf1">Hash-chained audit log</text>
<text x="240" y="494" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="10.5" fill="#94a0b0">grants, denials, rejections · tamper-evident</text>
<text x="240" y="552" text-anchor="middle" font-family="'JetBrains Mono',monospace" font-size="11.5" font-weight="600" fill="#5eead4">a token can only narrow · never widen</text>
</svg>
<div class="diagram-steps" aria-hidden="true">
<div class="dstep"><p class="dstep-title">Asha · user</p><p class="dstep-sub mono">read:calendar read:profile write:notes · 3600s</p></div>
<div class="darrow">↓ keeps read:calendar write:notes · drops read:profile</div>
<div class="dstep"><p class="dstep-title">planner-bot</p><p class="dstep-sub mono">aud: calendar-api · 600s</p></div>
<div class="darrow">↓ keeps read:calendar · drops write:notes</div>
<div class="dstep"><p class="dstep-title">summarizer-tool</p><p class="dstep-sub mono">nested delegation · 300s</p></div>
<div class="darrow">↓</div>
<div class="dstep deny"><p class="dstep-title">widening attempt</p><p class="dstep-sub mono">invalid_scope: admin:all denied</p></div>
<div class="darrow">↓</div>
<div class="dstep"><p class="dstep-title">key-bound token (cnf.jkt)</p><p class="dstep-sub mono">stolen token, no key: rejected</p></div>
<div class="darrow">↓</div>
<div class="dstep tint"><p class="dstep-title">Hash-chained audit log</p><p class="dstep-sub mono">grants, denials, rejections · tamper-evident</p></div>
<p class="dnote mono">a token can only narrow · never widen</p>
</div>
<figcaption>The demo's actual exchange story, rendered from real responses: scope and lifetime shrink at each hop, a widening request is denied, and a stolen token fails without the key it was bound to.</figcaption>
</figure>
<div class="try-row">
<a class="button button-primary" href="https://actas-rajatendu.vercel.app" target="_blank" rel="noopener">Try the live demo</a>
<p class="try-note">Build a delegation chain, then watch a widening attempt get refused, a copied token rejected and a replayed proof rejected - on a fresh throwaway instance. A policy workbench builds boundary cases, edits one agent's limits through a guided form, and compares a candidate against a saved baseline to show which decisions flip before you adopt it - a planning tool, not a running authorization server.</p>
</div>
<p class="detail-label mono">The engineering</p>
<ul class="feature-list">
<li>Agents receive tokens that can only shrink: less scope, bounded lifetime, narrowed audience, with the full delegation chain embedded in the token.</li>
<li>Per-agent policy at every hop, and every grant and refusal recorded in a hash-chained audit log.</li>
<li>A standalone resource server verifies tokens over remote JWKS in a separate process: the real topology, not shared state.</li>
<li>DPoP-style ES256 sender constraint (method, URL, time, token hash): at the token endpoint a live proof is verified and the issued token is bound to that key via a <code>cnf.jkt</code> claim, so presenting a token later requires the key it was bound to, and a replayed proof is rejected; on a persistent single instance the proof cache survives restarts, and an opt-in server-nonce mode adds the RFC 9449 challenge flow (120-second, single-use).</li>
<li>The original threat document has not been recovered; responses carry hardening headers (content-type sniffing disabled, a content security policy restricting connections to the same origin (inline scripts, styles and Google Fonts allowed), per-request IDs for tracing).</li>
<li>Operator tooling: structured JSON request logs (method, path, status and duration only, never headers, queries or bodies), per-process metrics with a request ID per call, and signing-key rotation with a retirement window.</li>
</ul>
<dl class="facts">
<div><dt>Verification</dt><dd>Recovered 0.14.1 baseline, checked 11 Nov 2026: 48 pytest cases (28 new recovery cases plus the 20-case lab regression suite), 17 integration checks, 26 process checks and desktop/390px browser workflows. These are new recovery checks, not the lost original development suite; original test and operational/threat files have not been recovered.</dd></div>
<div><dt>Stack</dt><dd>Python, FastAPI, RS256 JWTs, ES256 sender constraints, JWKS endpoint with key rotation</dd></div>
</dl>
<div class="limits">
<p class="limits-title">Honest limits</p>
<p>Not a production authorization server and not independently audited. The public demo is stateless and in-memory: every lab run uses a fresh throwaway instance that stores nothing, so nothing carries across runs, and admin and revocation endpoints are disabled there. Metrics on the public demo are per-instance and reset on restart. Sender constraints are a DPoP-style subset: replay tracking is per server, durable across restarts only when the server runs with persistence, so on a multi-instance host it is best-effort, not global replay prevention. The opt-in server-nonce mode (120-second, single-use) is off on the public demo. The hash-chained log is tamper-evident, not tamper-proof.</p>
</div>
<div class="project-links">
<a class="text-link" href="https://actas-rajatendu.vercel.app" target="_blank" rel="noopener">Open the live demo</a>
<p class="project-links-note">Build a delegation chain, narrow scope across hops, attempt a forbidden widening and bind a token to a key: the lab accepts the holder, rejects a copied token and rejects a replayed proof, all on a fresh throwaway instance that stores nothing. Actas also runs embedded inside llmgate as its delegation layer - the gateway demo runs a full exchange against simulated models.</p>
</div>
</article>
<article class="project-card reveal">
<div class="project-head">
<div>
<p class="project-eyebrow mono">Project 03 · Simulation, labelled as simulation</p>
<h2>paper-trading-india</h2>
</div>
</div>
<p class="project-lede">A public paper-trading experiment tracker for Indian equities, built to test a short-term momentum strategy under strict honest-reporting rules. The first experiment, EXP-001, replays three years of daily data on 50 current NIFTY 50 constituents with Rs 1,00,000 in virtual capital and an itemized delivery-cost model. Every decision is logged at generation time, and an independent 449-check script reproduces each published number exactly.</p>
<div class="metrics">
<div class="metric">
<p class="metric-value mono">+11.15 <span>%</span></p>
<p class="metric-label">Strategy, 3-year simulated window</p>
<p class="metric-sub mono">net of modeled costs · adjusted data</p>
</div>
<div class="metric">
<p class="metric-value mono">+27.8 <span>%</span></p>
<p class="metric-label">NIFTY 50 price return, same window</p>
<p class="metric-sub mono">gross · no costs · no dividends</p>
</div>
<div class="metric">
<p class="metric-value mono">449 <span>checks</span></p>
<p class="metric-label">Independent reproduction script</p>
<p class="metric-sub mono">0 failures on public main</p>
</div>
<div class="metric">
<p class="metric-value mono">96 <span>fills</span></p>
<p class="metric-label">86 more buys skipped, cash-limited</p>
<p class="metric-sub mono">next-session open · ±10 bps slippage</p>
</div>
</div>
<p class="metrics-note">EXP-001, 15 Nov 2023 - 13 Nov 2026, historical simulation on adjusted Yahoo daily data. The two returns use unmatched treatments (strategy net of modeled costs versus a gross price index), so the gap is not a measured alpha difference. Paper only - no real orders, ever.</p>
<p class="detail-label mono">The engineering</p>
<ul class="feature-list">
<li>Decisions are logged at generation time with run-date-suffixed ids so the record cannot be silently rewritten later; the one deterministic backfill from the first run is labelled as a backfill in the repository.</li>
<li>Execution happens one trading session after each decision, at that day's open plus or minus 10 bps slippage - never on the same bar that produced the signal.</li>
<li>Costs are itemized in one auditable module - brokerage, STT, exchange and SEBI charges, GST and stamp duty - and every fill's charges are re-verified by the reproduction script.</li>
</ul>
<dl class="facts">
<div><dt>Verification</dt><dd>EXP-001 published 14 Nov 2026: 31 pytest cases plus a 449-check independent reproduction of every recorded number (final equity Rs 1,11,150.78; maximum drawdown Rs 14,627.16; 96 fills; 86 cash-limited skips; benchmark +27.8%), all passing on the public main branch.</dd></div>
<div><dt>Stack</dt><dd>Python, yfinance daily data, pytest</dd></div>
</dl>
<div class="limits">
<p class="limits-title">Honest limits</p>
<p>A historical simulation, not investment evidence: the universe is today's constituents, so it is survivorship-biased; the cost model is an assumption itemized in the repository, not a quote; results come from a single run; and the benchmark comparison uses unmatched treatments. An out-of-sample forward mode exists locally but is unpublished pending review. Nothing here is investment advice, and no return or loss cap is promised.</p>
</div>
<div class="project-links">
<a class="text-link" href="https://github.com/Rajatendu1/paper-trading-india" target="_blank" rel="noopener">Open the repository</a>
<p class="project-links-note">No live demo: the project is a data experiment, and its results are reproducible from the public repository instead.</p>
</div>
</article>
</section>
<section class="section reveal" id="open-source">
<p class="section-label mono">Open Source Contributions</p>
<article>
<p class="project-kicker mono">gorakhargosh/watchdog · Merged 14 Nov 2026</p>
<div class="project-head"><h2>Debouncer tests that catch regressions</h2></div>
<p class="project-summary">Strengthened the <code>EventDebouncer</code> tests in the watchdog file-watching library. The old starvation test also passed when elapsed-time flushing was removed entirely; the rewritten tests synchronize the debouncer lifecycle and add starvation regressions that fail under that mutation.</p>
<p class="about-text">Test-only: one test file changed, no production code touched, and the PR does not claim the underlying bugs are still open. Accepted and merged upstream in watchdog.</p>
<a class="text-link" href="https://github.com/gorakhargosh/watchdog/pull/1458" target="_blank" rel="noopener">View merged PR #1458</a>
<p class="project-links-note">Tests prepared and executed by Instinct, an AI assistant, at my request.</p>
</article>
<article>
<p class="project-kicker mono">nushell/reedline · Merged 14 Nov 2026</p>
<div class="project-head"><h2>Completion highlights on the right letters</h2></div>
<p class="project-summary">Fixed fallback completion highlighting in reedline, the line editor behind Nushell. When a completer supplies no match indices, the fallback searched a lowercased copy and returned UTF-8 byte offsets, while the styler expects grapheme indices in the original text - underlining drifted after accented characters, emoji and case expansions. The fix maps those offsets back to the original displayed graphemes.</p>
<p class="about-text">Completer-supplied match indices and the public API are unchanged. Validated with 18 added cases plus the library suite and doctests run locally. Accepted and merged upstream in reedline.</p>
<a class="text-link" href="https://github.com/nushell/reedline/pull/1303" target="_blank" rel="noopener">View merged PR #1303</a>
<p class="project-links-note">Patch prepared and tests executed by Instinct, an AI assistant, at my request.</p>
</article>
<article>
<p class="project-kicker mono">sqlfluff/sqlfluff · Merged 15 Nov 2026</p>
<div class="project-head"><h2>TRANSFORM statements, lintable</h2></div>
<p class="project-summary">Added PostgreSQL <code>CREATE TRANSFORM</code> and <code>DROP TRANSFORM</code> parsing to SQLFluff. Scripts that define procedural-language type conversions - PostgreSQL's <code>hstore_plpython</code> extension, for example - previously failed with parse errors and now lint cleanly.</p>
<p class="about-text">Fixtures cover <code>OR REPLACE</code>, either or both conversion directions, optional signatures, schema-qualified and quoted names, <code>IF EXISTS</code> and <code>CASCADE</code>/<code>RESTRICT</code>, and all twelve new statements were validated against native PostgreSQL 18 grammar via pglast. Accepted and merged upstream in SQLFluff.</p>
<a class="text-link" href="https://github.com/sqlfluff/sqlfluff/pull/9140" target="_blank" rel="noopener">View merged PR #9140</a>
<p class="project-links-note">Patch prepared and tests executed by Instinct, an AI assistant, at my request.</p>
</article>
<article>
<p class="project-kicker mono">pyenv/pyenv · Merged 15 Nov 2026</p>
<div class="project-head"><h2>Custom header paths, actually honored</h2></div>
<p class="project-summary">Fixed custom dependency header detection in python-build, the build engine inside pyenv. Its <code>configured_with_package_dir</code> check only inspected global <code>CPPFLAGS</code> while processing a <code>CPPFLAGS=</code> configure argument, missed package-specific flags entirely, and read the package's make-option array instead of its configure-option array - so a custom Readline header path could be ignored and Homebrew's paths prepended ahead of the user's own.</p>
<p class="about-text">The fix checks global and package-specific <code>CPPFLAGS</code> independently against the configure options; the header-existence check and package-manager fallback are unchanged, and no shell-style quote parsing was added. Validated with 10 added regression tests (6 fail against unchanged upstream, all pass with the fix) plus the full suite at 547 passed, 10 skipped. Accepted and merged upstream in pyenv.</p>
<a class="text-link" href="https://github.com/pyenv/pyenv/pull/3633" target="_blank" rel="noopener">View merged PR #3633</a>
<p class="project-links-note">Patch prepared and tests executed by Instinct, an AI assistant, at my request.</p>
</article>
<article>
<p class="project-kicker mono">jgrapht/jgrapht · Merged 17 Nov 2026</p>
<div class="project-head"><h2>Biconnectivity without the stack overflow</h2></div>
<p class="project-summary">Replaced the recursive Hopcroft-Tarjan DFS behind JGraphT's biconnected components, cutpoints and bridges with explicit heap frames: deep graphs that overflowed the Java call stack (reported in issue #1266) now traverse in the same O(V + E) time, with the public API and lazy computation unchanged.</p>
<p class="about-text">Tests include 10,000-vertex paths and cycles that fail with StackOverflowError before the fix under -Xss1m and pass after it, all 1,024 simple graphs on five vertices checked against independent removal/BFS oracles, and 500 seeded pseudographs including parallel edges and self-loops. Accepted and merged upstream in JGraphT.</p>
<a class="text-link" href="https://github.com/jgrapht/jgrapht/pull/1417" target="_blank" rel="noopener">View merged PR #1417</a>
<p class="project-links-note">Patch prepared and tests executed by Instinct, an AI assistant, at my request.</p>
</article>
<p class="project-links-note">Highlights are capped at the five most recent merged contributions; open work and earlier entries stay in GitHub history. Research, patches and tests by Instinct, an AI assistant, at my request.</p>
</section>
<section class="section reveal" id="bug-reports">
<p class="section-label mono">Bug reports and investigations</p>
<p class="section-intro">Reported upstream, separate from merged fixes. All four reports were open and unassigned when checked on 17 Nov 2026. Local reproductions are not observed Microsoft Graph or Google Cloud incidents or maintainer-confirmed fixes.</p>
<div class="oss-ongoing">
<ul class="oss-ongoing-list">
<li><span class="oss-status">Open report · #900</span><a class="text-link" href="https://github.com/microsoft/kiota-python/issues/900" target="_blank" rel="noopener">microsoft/kiota-python: Redirect errors that exhaust a connection pool</a><span>Streaming responses left open on redirect-limit or request-construction errors blocked a later request in a one-connection localhost test. Nine regression/control cases: six fail before cleanup, all nine pass with the local candidate. Local candidate only; no PR submitted.</span></li>
<li><span class="oss-status">Open report · #901</span><a class="text-link" href="https://github.com/microsoft/kiota-python/issues/901" target="_blank" rel="noopener">microsoft/kiota-python: Detached models restored into a backing store</a><span>Stale nested-model subscriptions can restore replaced or cleared values; replacement edits can go untracked. This is a backing-store API reproduction, not demonstrated HTTP PATCH corruption. 18 focused cases: 16 fail on baseline, all 18 pass with the local candidate; 282 package tests pass. Local candidate only; no PR submitted.</span></li>
<li><span class="oss-status">Open report · #19103</span><a class="text-link" href="https://github.com/googleapis/google-cloud-python/issues/19103" target="_blank" rel="noopener">googleapis/google-cloud-python: Non-blocking refresh after credential serialization</a><span>A pickle roundtrip leaves OAuth2 credentials without the worker expected by non-blocking refresh, causing missing-worker exceptions for stale or expired tokens. Nine mocked local cases: four failures and five passing controls; no real token or network request. Reproduction only; no fix implementation or PR.</span></li>
<li><span class="oss-status">Open report · #19104</span><a class="text-link" href="https://github.com/googleapis/google-cloud-python/issues/19104" target="_blank" rel="noopener">googleapis/google-cloud-python: Async retry drops a recovery yield</a><span>AsyncStreamingRetry discards a value yielded while handling athrow(), skipping recovery output and disrupting the following send. Eight local cases: three failures and five passing controls on Python 3.10 and 3.13; no Cloud requests. Reproduction only; no fix implementation or PR.</span></li>
</ul>
<p class="project-links-note">Investigation, candidate patches and reproduction tests developed with Instinct, my AI assistant, at my direction. Each report discloses AI assistance.</p>
</div>
</section>
<section class="section reveal" id="pipeline">
<p class="section-label mono"><span class="section-number">02</span> The build pipeline</p>
<p class="section-intro">Next entries in the series. Each is picked to exercise skills that current backend, platform and AI-infrastructure openings actually list, and each ships tested, measured and written up.</p>
<div class="pipeline-list">
<div class="pipeline-item">
<p class="pipeline-number mono">04</p>
<div class="pipeline-body">
<p class="pipeline-name">Postgres realtime engine</p>
<p class="pipeline-detail">Stream row changes to connected clients using logical replication and row-level security.</p>
</div>
<p class="pipeline-status mono">Candidate</p>
</div>
</div>
</section>
<section id="about" class="section reveal">
<p class="section-label mono"><span class="section-number">03</span> About</p>
<p class="about-text">Most recently I was SDE II on the AWS UnifiedAuth platform in Hyderabad: service-to-service identity, authorization and trust propagation for agentic workflows. Before that, Spring Boot backend for HSBC's UK and Hong Kong mobile apps, preceded by a machine learning internship at HighRadius.</p>
<dl class="facts">
<div><dt>Core stack</dt><dd>Python, Java, Spring Boot; AWS, Kubernetes, Docker, CI/CD</dd></div>
<div><dt>Focus areas</dt><dd>Backend platforms, auth and identity (OAuth2), AI agents, observability</dd></div>
<div><dt>ML background</dt><dd>Time series, anomaly detection, forecasting (Kats, TensorFlow, scikit-learn)</dd></div>
</dl>
</section>
<section id="contact" class="section section-contact reveal">
<p class="section-label mono"><span class="section-number">04</span> Contact</p>
<h2 class="contact-heading">Let's talk.</h2>
<div class="contact-links">
<a class="text-link" href="mailto:rajatendud@gmail.com">rajatendud@gmail.com</a>
<a class="text-link" href="https://linkedin.com/in/rajatendu-dey" target="_blank" rel="noopener">linkedin.com/in/rajatendu-dey</a>
<a class="text-link" href="https://github.com/Rajatendu1" target="_blank" rel="noopener">github.com/Rajatendu1</a>
</div>
</section>
</main>
<footer class="footer">
<p>Portfolio opened November 2026, content updated 17 Nov 2026. llmgate and actas facts verified against their tested builds and live deployments on 11 Nov 2026; paper-trading-india facts verified against its public repository on 14 Nov 2026. New entries land when a project reaches a tested, demo-able state.</p>
</footer>
<script>
(function(){
if(!('IntersectionObserver' in window)){return;}
document.documentElement.classList.add('js');
var els=document.querySelectorAll('.reveal');
if(window.matchMedia('(prefers-reduced-motion: reduce)').matches){
els.forEach(function(e){e.classList.add('in');});return;
}
var io=new IntersectionObserver(function(entries){
entries.forEach(function(en){
if(en.isIntersecting){en.target.classList.add('in');io.unobserve(en.target);}
});
},{threshold:0,rootMargin:'0px 0px -12% 0px'});
els.forEach(function(e){io.observe(e);});
})();
</script>
</body>
</html>