B.Tech Computer Science & Engineering @ IMS Engineering College (AKTU) · BS Computer Science @ University of the People
I build programming languages, CPUs, and operating systems from scratch — and small language models on top of them.
Right now: shipping Qasida v2.0.0 (on PyPI) and self-hosting its compiler — the lexer and parser are done and verified, the type checker is in progress.
| Project | What it is |
|---|---|
| Qasida | A safe, simple, natively fast language built from scratch — v2.0.0, on PyPI (pip install qasida), native/WASM/big32 backends, freestanding mode |
| big32 | A 32-bit RISC-V CPU built from zero in Verilog, booting a Linux kernel into early driver init |
| Provenance-Aware Belief Aggregation | A 17-page research manuscript on multi-agent communication |
| Black Moon OS | Arch-based Linux distro with a macOS-style look, plus JalOS, a teaching OS in Qasida — bootable ISO built Oct 2026 |
| OpenMythos | A miniature recurrent language model, built from scratch — 44M trained on TinyStories, 1B roadmap underway |
| Nyx | A decision-intelligence simulator with a live demo |
| ASIC Puzzle 2026 | A chip reverse-engineered from its GDS layout alone |
| Recurrent Small LMs | NeurIPS-style technical report (22 pages) on recurrent small language models |
| NYX Phoenix Migration | Enterprise multi-agent simulation platform — NYX migrated from Streamlit to Next.js + FastAPI, same deterministic kernel, zero logic loss |
| TINY8 | An 8-bit CPU built from zero in Verilog — sum 1..10 = 55, verified two ways |
Qasida — a programming language built from scratch: static types, no null, immutable by default, checked arithmetic, native/WASM/RISC-V backends, and a freestanding mode on the road to a from-scratch OS. v2.0.0, on PyPI.
big32 — a 32-bit RISC-V CPU built from zero in Verilog, now booting a Linux kernel (6.6.52) into early driver init. SBI firmware, device tree, CLINT, PLIC, virtio block, packet networking — a full SoC, not just a core.
Provenance-Aware Belief Aggregation — when agents report through different communication conventions, weighting messages by source reliability answers the wrong question. This work proves structural limits of scalar trust weighting and introduces a Bayesian aggregator that decodes messages through inferred latent conventions. 17-page manuscript, headline runs over 190 seeds. Beginner explainer
OpenMythos — a from-scratch recurrent language model: does weight-shared recurrent iteration buy effective depth more cheaply than wider layers? 44M on TinyStories (one T4): loop depth 4 loss 1.484 vs depth-1 baseline 3.380 — but depth 8 is worse, so extrapolation is scale- and task-dependent. 1B config (1.06B params) ready. 44M report
Nyx — Decision Intelligence Simulator — give it a hard decision and a miniature society of AI agents debates it, with a deterministic kernel (same seed, same outcome) and a strategic forecast of risks and influencers. Live demo
Jane Street ASIC Puzzle 2026 — solved — reverse-engineered a chip from its GDS layout alone, by four independent methods that all agreed. Serial-key lock, 121-bit key, flag (* TWO STARS *). Published after submissions closed; nothing was submitted.
Plain-language explainers, one analogy per project, with verified vs reported vs plan-only labels: PVI · Nyx · OpenMythos 44M · MiniMythos 125M · 1B roadmap · ASIC puzzle
- BS in Computer Science — University of the People (WSCUC-accredited) — Student ID: C211425106
- B.Tech in Computer Science & Engineering — IMS Engineering College (AKTU)
Python · PyTorch · C · Verilog · NumPy · LaTeX · Git · RISC-V
Interests: programming languages · compilers · operating systems · small language models · multi-agent systems · Bayesian inference · hardware
📫 Reach me: dubeymaya787@gmail.com

