I build AI systems and the engineering underneath them — machine learning, LLM agents, and the infrastructure that keeps both honest.
Early in my career, and I'd rather say that plainly than pad it. No job titles yet — just a habit of building the thing instead of reading about it.
| Building | Learning | Exploring |
|---|---|---|
| AI-powered applications | Machine learning | LLMs · Agents |
| Developer tools | Data structures & algorithms | Computer vision |
| Research-style experiments | AI systems design | AI infrastructure |
Where the work happens — the projects themselves, not a PR tally.
| Ecosystem | Domain |
|---|---|
| Kubeflow | ML workloads on Kubernetes |
| PyTorch Lightning | Training loops that scale without the boilerplate |
| DeepMind Concordia | Generative agent simulation |
| Langflow · OpenRAG | Visual pipelines for LLM and RAG systems |
| Open Climate Fix | Solar forecasting to cut grid emissions |
| Microsoft AzureTRE | Trusted research environments on Azure |
| Django | The Python web framework for perfectionists |
| Google etils | Shared Python tooling used across Google |
Axon — knowledge infrastructure that tries to stay true. Most internal AI tools are a chat box bolted onto a vector store: the moment the code moves, they answer confidently and wrongly, and nothing in the system knows it has gone stale. Axon models knowledge as atomic, attributable claims instead of loose embeddings, then continuously checks those claims against their sources. Drift detection flags knowledge whose source changed underneath it.
ingest → extract claims → map → detect drift → maintain truth
Python FastAPI PostgreSQL pgvector Next.js TypeScript — Active. A concept under development, not a released product.
Also built
- Email Agent Environment — email as an explicit task / state / action environment, so every step of an agent's reasoning is inspectable instead of one long prompt and a hope.
Experimental - Resource Management System — academic material lives across a dozen group chats; this puts it behind one searchable schema.
Built - Sundarone — a café & resort platform built as a real application: typed content layer, installable, fast on a weak connection.
In progress
Repo links are omitted where the code isn't public yet — a dead link is worse than none.
Bold is what I reach for first.
| AI / ML | Python · PyTorch · scikit-learn · Hugging Face · LLMs · Embeddings · NumPy · Pandas |
| Backend | FastAPI · Node.js · REST APIs · Django |
| Data | PostgreSQL · pgvector · MySQL |
| Frontend | React · Next.js · TypeScript · Tailwind · Angular |
| Tooling | Git · Docker · Jupyter · Vercel · Render |
Baselines first — cosine similarity before a fine-tune, Postgres before a vector store. Then build it, break it, and go again.
Open to collaboration, open-source work, and conversations about ML systems.
portfolio · goyalvishal7711@gmail.com · LinkedIn · GitHub
Counters above come straight from the GitHub API. Everything else here is something I built or somewhere I contributed.