AI systems, agents and developer tooling. 19, B.Tech CSE (AI & Data Science) student in Pune, India. Open to work.
I fix real bugs in AI-agent and developer tooling, then write down what I verified and what I did not.
| Project | What it is | Honest status |
|---|---|---|
| AIOS | Python core that turns a goal into a plan, runs it, and verifies the result. File and directory checks look at the real disk instead of trusting the agent's report. Run python examples/verified_delegation.py to see an agent's false "done" get caught. |
Early beta. Real execution covers terminal steps only. No CI yet. The desktop app is not fully tested. |
| Anthropic performance take-home | My solution to Anthropic's public kernel-optimization take-home: about 147K down to about 14.4K simulated cycles (10.2x) | Did not reach the benchmark's tighter cycle thresholds; the repo says so |
| micrograd-JAX | Karpathy's micrograd idea rebuilt with JAX transforms (grad, jit, vmap) as a learning project |
Fork of the original, my JAX version is in the notebook |
| ML Fundamentals | 19 notebooks on core ML algorithms | Learning notes, not a library |
| agent-canvas | Typed Python DSL for visual lessons, with a test suite (312 tests passed in local verification; requires Python 3.12+) | Locally verified; not a deployment claim |
Merged
- E2B #1912: stopping a filesystem watch surfaced as a timeout error
- Meta Muse gadget SDK #3: a command could hang after its timeout because a detached process held the output pipes open
- AgentPhone MCP #74: the documented
--portflag was ignored - tester-army/e2e #734: migration docs said
toContainmatched Playwright when it did not
Fixes adopted or credited by maintainers
- gstack #3001 and #2997: incorporated into the maintainer's merged PR #3013
- AlphaFold 3: invalid chain IDs passed validation and failed later; the maintainer commit credits the report
Open for review
- Hugging Face tau #762, Composio #4700, llama-cookbook #1081, fvcore #159, TensorDict #1839, Ollama #18745, AgentPhone MCP #79
Tally: 34 pull requests opened in other people's repositories as of 6 Oct 2026 (full list): 4 merged, 20 open, 10 closed without merge (including the 2 gstack PRs whose fixes were incorporated). Forks and my own repos are not counted. Snapshot updated by hand.
PTF: Protection Toward Future. An empirical incident study of autonomous AI agent behaviour on a personal workstation, with a proposal for infrastructure-level safety architecture. Read it on SSRN.
I'm looking for full-time or internship roles in AI systems, agents and developer tooling, remote or India-based. I'm a student, and I'm flexible about how the right role fits around my degree.