CS Undergrad @ BITS Pilani | AI/ML Developer | GenAI β’ Agentic AI β’ Deep Learning β’ LLM Applications
π 18 PRs merged upstream β steipete/CodexBar (19k+ β, 11) Β· mem0ai/mem0 (62k+ β, 2) Β· huggingface/OpenEnv (2.4k+ β, 1) Β· +4 more. Each one found by reading the code, reproduced with a failing test, then fixed. See the list β
auto-refreshed daily via GitHub Actions β real repo/star/follower/commit counts, no fake numbers
Currently:
- π Building DeepLense-AI-Scientist β multi-agent framework orchestrating scientific workflows in gravitational lensing research (Pydantic AI)
- π Contributor on steipete/CodexBar (19k+ β) β 11 PRs merged, incl. a macOS Keychain fix (#2102) and the
codexbar guardCLI (#2237) - π§© Shipping fixes into mem0ai/mem0 (62k+ β), huggingface/OpenEnv and sktime/sktime β scope-isolation, vector-store scoring, and client API bugs, each with redβgreen regression tests
- βοΈ Invited by co-founder Nikhil Pareek to contribute to future-agi/future-agi β 4 PRs open on the LLM eval & observability platform
| Project | What it does |
|---|---|
| bah2026-ps14 | ISRO BAH 2026 PS14 β forecasts >2 MeV electron flux at GEO orbit 30β45min/6h/12h ahead via a 5-stage pipeline (persistence β RandomForest β LSTM/Transformer), validated against ISRO GRASP/GSAT |
| Proxim | Real-time AI HCP roleplay avatar trainer for pharma sales reps β live voice sessions, 3D avatar, post-call coaching scorecard |
| SQL-Query-Optimization-Environment | RL environment where agents rewrite slow SQL, graded on real DuckDB execution timing + correctness across 5 anti-pattern task types |
| et-hackathon-ps7 | AI platform detecting behavioral anomalies in network/host telemetry β not signature-based |
β CodexBar (19k+ stars) β macOS menu bar app for AI usage tracking, by Peter Steinberger (creator of OpenClaw, ex-PSPDFKit founder) Β· Contributor, 11 PRs merged
- #2102 fix: Claude no-prompt Keychain repair for missing credentials file β merged, closes #1975. Diagnosed a Keychain-access gate that blocked even guaranteed no-UI reads under "Avoid Keychain Prompts," shipped the fix with regression tests + a live macOS verification script, reviewed and merged by the maintainer.
- #2237 feat:
codexbar guardβ quota-aware exit code to gate automation β merged. Self-proposed feature: a CLI subcommand that exits non-zero when a provider's quota is spent, so scripts and CI can stop before burning a rate limit. - #2483 fix: validate LiteLLM and LLM Proxy base URLs before sending the API key β merged. A malformed or attacker-supplied base URL was accepted as-is, so the configured key went out with the request; the URL is now validated first.
- Linux availability β merged: #2475 (MiniMax with a configured API key), #2356 (Alibaba/Qwen Token Plan with a manual cookie). Both providers were gated behind a macOS-only web-auth path even when the user had already supplied a credential the CLI could use.
- Usage-percent correctness series β merged: #2255 (Cursor), #2265 (Abacus), #2293 (ElevenLabs), #2342 (z.ai). Providers reporting over-quota or already-scaled values rendered as impossible percentages in the menu bar; each fix landed with a redβgreen test transcript in the PR body.
- Reset-countdown fixes β merged: #2343 (a countdown of exactly 24h rendered as "24h" instead of rolling over to the day form, across four formatters), #2335 (LLMProxy picked an already-elapsed timestamp as the next reset).
- OpenCode/OpenCodeGo percent unit double-scaling β a fraction heuristic also ran on the already-computed
used/limitpercent, so real usage under 1% got multiplied by 100. Fix landed in #2331 (commit authored by me).
β mem0 (62k+ stars) β universal memory layer for AI agents, by Mem0 (Y Combinator S24)
- #6343 fix(ts-oss): don't let
update()metadata overwrite user_id/agent_id/run_id β merged. Caller metadata was spread over the scope identifiers, silently re-homing a memory to a different user/agent/run. - #6435 fix(vector-stores/baidu): convert L2 distance to similarity score in
search()β merged. Baidu VectorDB returned raw L2 distance where every other store returns a similarity, inverting relevance order and breaking score thresholds. - Open: reranker candidate-pool fix (#6449 Python / #6537 TS), a hash-dedup TOCTOU race in
add()(#6516 Python / #6532 TS), Upstash metadata filter-injection escaping (#6610),add()metadata setting a memory's identity scope (#6656), and two provider-config fixes (#6302, #6257).
β OpenEnv (Hugging Face, 2.4k+ stars) β environment interface library for RL post-training
- #959 feat(env-client): sync bootstrap constructors + public
base_urlβ merged, closes #935. Madefrom_docker_image/from_hubusable from sync code and exposed the resolved base URL instead of forcing callers into private attributes. - Open: #1008 path traversal via agent-supplied names in
finqa_env, #1011 discovery cache moved out of world-writable/tmp, #1006 execution-grounded DuckDB SQL-optimization environment.
- π Merged PR #2475 in steipete/CodexBar
- π Merged PR #2483 in steipete/CodexBar
- π£ Commented on #2475 in steipete/CodexBar
- π£ Commented on #2483 in steipete/CodexBar
- πͺ Opened PR #6656 in mem0ai/mem0



