Anuj
Attri
Software engineer. The products happen to be AI products.
Contents
- About2
- Journey3
- Toolkit4
- Full-Stack AI Workspace5
- Desktop Research Tool7
- ocalc9
- ggml / llama.cpp11
- Also13
Scroll, use the arrow keys, swipe, or click a page edge to turn. The tabs on the right jump to a section.
Software engineer: FastAPI and PostgreSQL on the server, React and TypeScript on the client, C++ where it has to be fast, and Docker, pytest and CI around all of it.
The products happen to be AI products. This book holds three projects and an open-source contribution, one spread each. A few smaller things are at the back.
Reachable by email.
Recognition
- Mitacs Globalink Research Scholarship2022
- Merit Scholarship, NYU Tandon2023
- Google beta tester, Jules and Flowpre-release
- Founder, TAM-VIT300+ members
Full stackSystemsAI
- Languages
- Python, C++, TypeScript, SQL
- Frameworks
- FastAPI, React, Electron
- Data
- PostgreSQL, SQLite, pgvector
- Delivery
- Docker, pytest, CI/CD
- AI
- PyTorch, ONNX, llama.cpp
01
Full-Stack AI Workspace
open source
A React and TypeScript client of 47 components on Vite, over a FastAPI backend that streams NDJSON. Guest mode serves it as a public site with no sign-up, each tab an isolated session deleted on close. A spend cap hands turns to a local model, interrupted runs recover from SQLite, and GitHub Actions runs 362 pytest cases, type checks and a Docker boot test.
On the surface
- Tool-using assistant
- Cited document answers
- Embedding search
- Guest sessions
- React
- TypeScript
- FastAPI
- SQLite
- Docker
02
Desktop Research Tool
active development
Ask a question; it finds and reads sources and writes a summary where each claim is cited. Nineteen component suites in Vitest with Testing Library and jsdom, and every build gated on tsc project references, ESLint and Prettier before electron-builder produces Windows, macOS and Linux installers.
On the surface
- Workspaces
- Question graph
- Cited summaries
- Reading view
- Electron
- React
- Vitest
- SQLite
- Private, walkthrough on request
03
ocalc
on PyPI, pip install ocalc
Runs a Python function across growing inputs and writes a one-line estimate of its time and space complexity, with the worst and median timings and the measurement protocol, as a comment above the function. A browser demo reads the worst case for a whole file.
On the surface
- Measure from the CLI
- Annotate the source
- Browser demo
- MCP server
- Python
- pytest
- Sphinx
- MCP
Smaller, complete
- AetherA desktop drawing and diagramming canvas driven by mouse, hand gestures from a webcam, or voice, with an assistant that edits the canvas from a sentence.
- BeaconAn always-on personal assistant for a 2 GB Jetson Nano: web UI, local memory, voice, and wake-on-LAN for the desk PC.
- LLM VisualizerA small transformer running in the browser, showing the tokens and next-token probabilities as temperature and top-p change. No backend.
Index
Thanks for reading.