Gavita Harshit

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.

LinkedIn | Email | GitHub

---

Projects

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

---

Open source contributions

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 --port flag was ignored
• tester-army/e2e #734: migration docs said toContain matched 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.

---

Research

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.

---

Activity

---

Let's talk

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.

Reach me on LinkedIn or email.