Shubhankar Tripathy lonexreb

📍 Palo Alto · 🎥 CV infra + AI-native sports analytics @ Paloa Labs · 🤖 Agent-native engineering

Building agents that actually work — event-driven systems, RL from human feedback, and tools that turn messy real-world interfaces into clean APIs. Currently the most recent external contributor to NVIDIA's Alpamayo (31 consecutive PRs, 1 merged).

🏆 Merged Contributions

_{shipped upstream}

_{awaiting merge}

_{in active review}

#	PR	Repo	What
1	`alpamayo#73`	NVlabs/alpamayo	Replace `assert` with `ValueError` so input validation survives `python -O`
2	`openclaw#77446`	openclaw/openclaw	Pin container-side workspace and config dirs in `docker compose`
3	`openclaw#74638`	openclaw/openclaw	Accept `browser.tabCleanup` keys in zod schema
4	`warp#9558`	warpdotdev/warp	Align OSS `.desktop` `Exec` with packaged binary name
5	`warp#9563`	warpdotdev/warp	Attribute Alacritty/vte derivative code in two more files
6	`warp#9667`	warpdotdev/warp	Recognize Mistral Vibe as a CLI agent
7	`warp#9670`	warpdotdev/warp	Route `CLIAgent::Pi` to the default session listener
8	`Megatron-Bridge#3649`	NVIDIA-NeMo/Megatron-Bridge	Add NVIDIA copyright headers to four package `__init__.py` files

📊 2026 YTD

$ git log --since=2026-01-01 --shortstat | langstats

Python      ████████████████████████████   73.6%
TypeScript  ██████                         16.6%
HTML        ██                              5.1%
Other       █                               3.5%
Rust                                        1.2%
─────────────────────────────────────────────────
1,023 commits  ·  5 languages  ·  30 active repos

🔨 What I'm Building

Developer Tools & Platforms

🌐 site2cli — Any website → CLI/API for AI agents. v0.6.0 on PyPI, MCP-native, 500+ tests.
🔬 MCPstudio — The Postman for MCP. Visual create/test/manage MCP servers, FastAPI + React + DDD.
🧩 zigsaw — AI workflow automation: chat or drag-and-drop, multi-LLM, instant API deploy, 100+ integrations.
🔭 kalam — Terminal agent for science / chip design / hardware / sec — arXiv, Yosys, ngspice, nmap, Ghidra MCPs. Named for Dr. APJ Abdul Kalam.
🔄 mobius — High-performance autonomous ML experimentation framework. Rust + Go + Python.
🌟 morningstar — Notion PRD → working code with tests + Slack updates. Powered by Claude Code.
🧱 deepagents — Batteries-included LangChain + LangGraph harness. pip install deepagents.
🐟 aish — Anthropic-marketplace Claude Code plugin — GPU control plane via TensorDock + Modal MCP.
📊 compgit — GitHub commits, quietly visible. Chrome / iPhone / macOS.

AI Agent Systems & RL

🐙 tentalis — ADHR meta-RL: agents that learn from manager feedback. GRPO / CISPO / DAPO + NATS + HaluGate.
⚡ RLYX-enhancer — Hackable GRPO with high-speed weight sync for multi-node RL.
🤖 AGI-INC — HALO-Agent for AGI Inc REAL Benchmark (112 web tasks, 11 sites). Gemini 3.1 Pro, 38.5% on zero-site.
🎓 university-sim — Multi-agent university benchmark: Concordia GM + OpenClaw + emergence metrics.
🌸 flwr-frl-kit — Federated RL Kit for flower.ai.
📈 DeepFinRL-UMass — Hierarchical MARL trader with Director + analyst agents, LLM sentiment, PPO/CPPO/GRPO.
🔬 autoresearch — Single-GPU research agents for Karpathy's nanochat — Lévy flights, Metropolis-Hastings, hippocampal replay, golden-angle sampling.
🧮 claude-code-math-skills — 7-agent Generator–Critic loop for rigorous math; Lean 4 + SymPy verifiers, Putnam/IMO benchmarks.

F1 & Motorsport AI

🏎️ aero-agent — CFD-aware research assistant for F1 aerodynamicists; optimize every sim hour under FIA ATR caps.
🔍 sim-scout — Paper → benchmarked OpenFOAM sim in hours, not weeks.
🏁 f1-design-ai — Blender geometry + OpenFOAM CFD + ParaView, orchestrated by OpenClaw.

Founder Performance & Health

🧬 FounderAgent — Health-aware agent: reads biometrics → blocks calendar, sets DND, sends recovery protocols.
🏆 FounderLeague — Anti-hustle-culture leaderboards: founders compete on readiness, sleep, recovery.
🧠 GameDay-AMS — Cognitive Performance System for Founders — sports science meets knowledge work.
🔋 FocusFuel-API — Wearable data → cognitive performance predictions. 35 tests passing.
🩺 AgentKit-Health — Marketplace of pre-built health agents — Sleep Coach, Burnout Detector, Recovery Agent.
💪 grithub — Sports + healthcare with BioGears physiological sim. Hacklytics 2025 winner.

Applied AI & Data

🏀 paloa-claw — Personal basketball AI training assistant.
🏥 clinical-trials-api — ETL + REST API for ClinicalTrials.gov; PostgreSQL, live on Render.
📧 sdr-automation — SDR research + outreach pipeline. YC AI Hackathon, built on Zigsaw.
📺 zigsaw-labs — Auto-generates, tests, optimizes short-form ads across TikTok, Reels, Shorts.
📝 grammer-paraphrase-system — Grammar correction + paraphrasing, clean DDD architecture.
🏠 HandyHommieAI — Home appliance command center: upload manuals, ask questions, auto-call manufacturer support.
🎯 growthclaw — Founder-scouting outbound for Crowdstake AI. Product Hunt / Reddit / IH at ~$0.02/lead. Austin Hackathon.
🔲 kalam-placer — Chip macro placement optimizer — Partcl + HRT $20K competition. ibm01 benchmark.

Fun & EdTech

♟️ dolphine-chess — Multi-tenant SaaS chess academy platform.
🍁 Maple — Emotion-aware AI tutor — Hume EVI + DBRX/LanceDB RAG + Midnight blockchain integrity.
🃏 blackjack-for-toddler — Toddler Blackjack teaching numbers + addition. React PWA + Capacitor for iOS.
🃏 poker-for-toddler — Texas Hold'em for kids (4–9) with toy cars, gems, crowns. React PWA on Vercel.
🕉️ Mahamrutyunjay-Mantra-Counter — macOS app for tracking mantra chanting via voice detection. 1–108 visual progress.
🇮🇳 xplncert — Inventors walk you through NCERT — AI persona videos for Indian class 6–12 textbooks.

🌍 Open Source Contributions

Shipping fixes, tests, and ergonomics upstream across NVIDIA's open AV / RL / training stack, agentic tooling, and terminal infra — selected PRs:

🌟 Featured: NVlabs/alpamayo — NVIDIA's open foundational driving model (Physical AI / AV) (31 PRs)

The last 31 consecutive PRs to the repo are mine (1 merged, 30 in review) — a sustained sprint hardening NVIDIA's open foundational driving model across input validation, runtime/attention correctness, type annotations, evaluation metrics, regression tests, docstring-vs-code drift, and first-run onboarding.

Validation & input handling

PR	Status	What it does
`#73`	✅ merged	Replace `assert` with `ValueError` in `load_physical_aiavdataset()` so input validation survives `python -O`
`#74`	open	Actionable HuggingFace dataset-access errors — surface gated-repo / 401 / 403 instead of confusing `IndexError` (fixes #59, #61)
`#77`	open	Replace `assert` with `ValueError` across public-API input validation
`#84`	open	Replace mutable default list with `None` sentinel in `basic_collation_fn`
`#94`	open	Replace `assert`s with `ValueError` in `chat_template/conversation.py`

Runtime & compatibility

PR	Status	What it does
`#75`	open	Pin expert decoder to SDPA so the model loads when Flash-Attention 2 is globally enabled (fixes #52)
`#78`	open	Remove deprecated `local_dir_use_symlinks` arg from `snapshot_download()` call
`#80`	open	Preserve integer/bool tensor dtype in `to_device()` (fixes #36)
`#99`	open	Use `dim=` instead of `axis=` in `comfort_reward._within_bound` (PyTorch idiom)

Type annotations & API contracts

PR	Status	What it does
`#85`	open	Fix `get_label_mask` docstrings and `get_assistant_mask` return type
`#89`	open	Document exclusive-end `chunk_ids` range and broaden type annotation in `pai_utils`
`#92`	open	Fix `MetricRunner.run()` return-type annotation and document its side effect
`#95`	open	Fix `viz.py` type hints — optional waypoints + `rotate_90cc` annotations
`#96`	open	Type-fix `init_wandb` `key` parameter and `_save_wandb_id` return type

New evaluation metrics

PR	Status	What it does
`#101`	open	Add Final Displacement Error (FDE) and minFDE metrics to evaluation suite
`#102`	open	Add trajectory smoothness metrics for evaluation

Docs & docstring sync

PR	Status	What it does
`#76`	open	Clarify coordinate-frame conventions for `project_waypoints_ftheta` (refs #34)
`#79`	open	Expand Troubleshooting with HF auth, FA2 `cu_seqlens`, and smoke tests for first-time users
`#82`	open	Sync README project-structure tree with on-disk layout
`#83`	open	Sync `load_physical_aiavdataset` docstring with the code
`#86`	open	Sync metric docstrings with the keys actually returned
`#87`	open	Document exclusive-end `--chunk` range in `curate_pai_samples.py`
`#93`	open	Sync `DistanceMetrics` docstrings with the actual signature and returns
`#98`	open	Fix `launch_alpamayo_model` docstring — checkpoint source is `--config` arg, not env var

Tests & DX

PR	Status	What it does
`#81`	open	Accept multiple `--chunk-ids` without quoting in `download_pai.py`
`#88`	open	Remove commented-out debug print in `QwenProcessor._preprocess_data`
`#90`	open	Add regression tests for `basic_collation_fn` and `get_assistant_mask`
`#91`	open	Add regression tests for `compute_minade` and `summarize_metric` keys
`#97`	open	Remove dead misplaced shebang from `convert_release_config_to_training.py`
`#100`	open	Add CLI flags to `test_inference.py` without changing default behavior
`#103`	open	Add `python -m alpamayo_r1.healthcheck` install smoke test

NVIDIA-NeMo/RL — NVIDIA's RL framework for LLM post-training (GRPO, DAPO, Megatron) (9 PRs)

Production-grade error handling, configuration ergonomics, and test coverage for the GRPO / Megatron training loop.

PR	Status	What it does
`#2395`	open	Clearer error when GRPO `overlong_filtering` finds no `truncated` field
`#2394`	open	Hard-fail when `nemo_gym` rollout collection returns fewer rows than requested
`#2393`	open	Extend worker offload guard to v2 + Megatron paths
`#2392`	open	Clear error when DTensor `train` / `get_logprobs` / `score` is called while offloaded
`#2391`	open	Expose Megatron checkpoint parallelism and RNG knobs to user config
`#2390`	open	Expose hardcoded Megatron infrastructure params to user config
`#2389`	open	Add author field to README citation BibTeX
`#2388`	open	Bump `accelerate` floor to 1.1.0 for `transformers` 5.3.0 compat
`#2387`	open	Test coverage for converter CLI entry points

NVIDIA-NeMo/Megatron-Bridge — NVIDIA's bridge between Megatron-LM and NeMo training recipes (9 PRs · 1 merged · 3 approved)

1 merged, 3 maintainer-approved — test coverage and observability across VLM / audio recipes, plus data-pipeline and TB-memory-aggregation fixes.

PR	Status	What it does
`#3649`	✅ merged	Add NVIDIA copyright headers to four package `__init__.py` files
`#3647`	✅ approved	Forward MoE / MTP metrics to MLFlow and Comet
`#3646`	✅ approved	Unit tests for `qwen2_audio` and `kimi_k25_vl` recipes
`#3645`	✅ approved	Aggregate TensorBoard memory metrics across PP group
`#3680`	open	Unit tests for HuggingFace dataset processors
`#3666`	open	Unit tests for `recipes/common.py` base helpers
`#3665`	open	Pad chat tensors and `loss_mask` in `pre_pad_dataset` (#2610)
`#3652`	open	Unit tests for `CometPlugin`
`#3650`	open	Unit tests for `pg_utils` helpers

openclaw/openclaw — agentic coding platform (18 PRs · 2 merged)

📝 Workflow notes & lessons learned from this contribution sprint →

PR	Status	What it does
`#77446`	✅ merged	Pin container-side workspace and config dirs in `docker compose`
`#74638`	✅ merged	Accept `browser.tabCleanup` keys in zod schema
`#77367`	open	Scope Discord `command-deploy` cache by application id (multi-tenant isolation, 3 regression tests)
`#75476`	open	Honor `model.compat.unsupportedToolSchemaKeywords` for OpenAI-completions tool schemas
`#75445`	open	Tolerate unresolved `SecretRef` tokens during Discord/Telegram channel-actions discovery
`#75339`	open	Normalize structured `delta.content` blocks to prevent `[object Object]` in chat replies
`#75248`	open	Reorder workspace `AGENTS.md` template so load-bearing rules come first
`#75217`	open	Honor `skipBootstrap` at the runtime injection path
`#74945`	open	Canonicalize `--model` to lowercase before dispatch
`#74891`	open	Skip API-key prompt when user skipped installing the skill
`#74643`	open	Per-agent `verboseDefault` and `elevatedDefault` config overrides
`#77184`	open	Re-export `StatusSummary`, `SessionStatus`, `HeartbeatStatus` from plugin-sdk
`#77189`	reviewed	Downgrade expected 1013 "gateway starting" close-before-connect log to debug
`#77191`	reviewed	Optional `agentId` filter on `cron.list` with default-agent fallback
`#77197`	open	Surface sharp-install hint when image optimizer exhausts every resize attempt
`#77215`	reviewed	Document Realtime Talk requires OpenAI Platform credits, not Codex subscription
`#77217`	open	Reserve `safe-area-inset-bottom` on sticky chat compose for iOS PWA
`#77224`	reviewed	Document BlueBubbles channel-vs-plugin disablement layers + safe loopback config

warpdotdev/warp — agentic terminal (18 PRs · 4 merged)