I build LLM systems where the model proposes and deterministic code decides.
AI Agents · LLM Infrastructure · Deep Learning
Open to full-time SWE / AI / ML Engineer roles.
14 merged pull requests across the ML stack — PyTorch · Anthropic · LangChain · vLLM · SGLang · Nous Research · LiteLLM — plus a first-author paper, an agent that designs real circuits, and products live in production.
| The fact | Why it's hard | |
|---|---|---|
| 🔥 | Merged into PyTorch core, reviewed by the TorchInductor lead (e9cfafa) |
Couldn't build torch locally — proved runtime equivalence by diffing both versions across 18 behavioral dimensions; then unblocked my own merge by proving an unrelated ROCm CI failure independent with four reproducible lines of evidence |
| ⚡ | An agent loop that designs real analog circuits (Summer 2026, Halo Microelectronics) | No numeric optimizer — LLM proposes experiment batches, a simulator holds authority over truth: 4 rounds, ~100× error reduction, 42 evaluations, 5 LLM calls, ~7 minutes, beating the accuracy floor in the company's own codebase |
| 🔬 | CUDA kernel correctness fix in vLLM core (~85k★) (#45466) | The issue thread blamed FlexAttention, CUDA graphs, and drivers — the real cause was an unchecked destination-pointer alignment in the KV-cache write path; fixed for every caller |
| 📡 | 77 kHz BLE RSSI firmware (Zephyr RTOS, nRF54L15) → 0.986 R² recovering signals 5.3× below Nyquist | 3× the highest published sampling rate, feeding a physics-informed network that recovers what classical sampling theory says is unrecoverable |
| 🧪 | RepoAgentBench — contamination-free coding-agent benchmarks (PyPI) | Public benchmarks overestimate agents by 20–50% via training-data contamination; mine freshly merged PRs so tasks postdate any model's cutoff — surfaced the same model producing opposite outcomes under two different harnesses |
| 🧵 | Speculative-decoding correctness fix in vLLM (#45352) | A silent config-propagation bug made a draft model instantiate at full 675B-scale dimensions in CI; fix required composing callable overrides and solving the pickling failure the composition introduced |
| 🏗️ | Multi-tenant cache-routing crash fix in SGLang (~29k★) (#26971) | A per-request key wasn't indexed per sub-request, silently collapsing prefix-cache isolation between tenants — fixed with a 6-path regression test across CPU, AMD, and CUDA CI |
| 💸 | ~6× prompt-token cost cut in LangChain (#1085) | Repo-wide static analysis found ensure_ascii defaults silently escaping CJK/emoji to \uXXXX across 11 sites in 3 modules — the kind of bug nobody sees because every request still succeeds |
| 🚢 | Shipped and live: archiagents.com · Chrome Web Store · LLM gateway | Real users, real uptime, real billing — including the honestly-reported finding that a rule-based heuristic beat Q-learning, so the heuristic shipped |
For recruiters · researchers · founders — 10-second version
- Recruiters: 14 merged PRs across PyTorch / Anthropic / LangChain / vLLM / SGLang · shipped products · Dec 2026 grad seeking SWE/AI/ML roles.
- Researchers: first-author on agent memory verification · contamination-free agent benchmarking · AAAI & IEEE-HKN member.
- Founders: I build end to end and ship — agent systems with real verifiers, a live AI product, and a production LLM gateway serving my own downstream apps.
Languages
ML Systems & AI
Backend & Web
Infrastructure & HPC
Embedded & Hardware
pytorch/pytorch — the framework everything else is built on
- PR #191866 — landed on
mainase9cfafa, reviewed and approved by @jansel (TorchInductor lead). (PyTorch merges viapytorchmergebot, which closes the PR once the commit lands — the commit link is the canonical record.) Converted three bare expressions in the compiler runtime that looked useless but carried side effects into explicit_ = exprbindings, and removed the genuinely dead statements intorch/fx/experimental/unification. The bare statements turned out to be a historical trick for suppressing F811 warnings; replaced with explicit# noqa: F811after establishing via minimal repro that ruff exempts underscore-prefixed names from F811 — undocumented behavior that explained why only the public names needed suppression. Unable to build torch locally, I proved runtime equivalence by importing the pre- and post-change packages side by side and diffing 18 dimensions: per-function behavior, exception types and messages, and the full dispatch registry.- Unblocking the merge: landing was blocked by a ROCm/gfx950 CI failure. I diagnosed it as
hipErrorIllegalState(HIP 401) thrown fromhipModuleLaunchKerneland proved it independent of my change with four reproducible lines of evidence — bytecode comparison, config-gating analysis (the relevant flags appear 0 times in the failure logs), the module never being imported, and elimination of cache bypass — and recommended the maintainer use@pytorchbot merge -i. The PR landed.
- Unblocking the merge: landing was blocked by a ROCm/gfx950 CI failure. I diagnosed it as
anthropics/claude-code-action (8.4k★) — Anthropic's official GitHub Action
- PR #1488 (merged): Closed a gap in the content sanitizer, which stripped injected instructions from inline images
but not reference-style images![alt][ref]. Added regression tests (771 passing); reviewed and merged intomainby an Anthropic engineer.
langchain-ai/langchain-aws — AWS/Bedrock integrations for LangChain
- PR #1085 (merged): Repo-wide static analysis caught
ensure_ascii=Truedefaults injson.dumpsacross Bedrock converters, tool-schema serializers, and stream parsers — silently escaping CJK/emoji to\uXXXXand inflating prompt token cost ~6×. Fixed across 11 sites in 3 modules.
vllm-project/vllm (~85k★) — the core LLM inference engine
- PR #45466 (merged): CUDA kernel correctness fix. Root-caused a
CUDA error: misaligned addresscrash (surfacing via FlexAttention withhead_size=46) that had been misattributed across the issue thread to FlexAttention, CUDA graphs, and GPU drivers. Real cause: the sharedvectorize_with_alignmenthelper only checked the input pointer's alignment — but inreshape_and_cache_flashthe destination KV-cache row isn't 16-byte-aligned for head sizes not a multiple of 8, so the kernel's 16-byte vectorized stores faulted. Added an output-pointer alignment check + scalar fallback, eliminating the unguarded-store hazard for every caller (incl. fp8/int8 quant kernels), Linux behavior byte-for-byte unchanged. Added a GPU regression test (head_size=46); merged into main by a core committer. - PR #45352 (merged): Speculative-decoding correctness fix. Root-caused a recurring CI OOM to a silent config-propagation bug: the draft-model config hardcoded its own
hf_overrides, silently dropping the target model's — so test-shrinking overrides never reached the Eagle draft, which instantiated at full 675B-scale dimensions. Fixed by composing the target's callable override with the draft's; also resolved a multiprocessing-pickling failure the composition introduced (nested closure →functools.partialon a static method) since vLLM's engine core pickles configs acrossspawn. Added a picklability regression test; re-enabled a previously-excluded test path. Shepherded and merged by a core maintainer.
sgl-project/sglang (~29k★) — high-performance LLM/multimodal inference-serving framework
- PR #26971 (merged): Fixed a batched multi-tenant cache-routing crash —
GenerateReqInput.extra_keywasn't indexed per sub-request, so the whole list was passed toRadixKey.child_key(), crashing prefix-cache matching withTypeError: unhashable type: 'list'. Added_normalize_extra_key()(scalar broadcast / list-length validation / parallel-sample expansion) + a 6-path regression test; passed 121 CI checks. - PR #25975 (merged, co-author): Prefill-delayer monitoring-metric fix —
prefill_delayer_wait_*histogram stuck at 0 because the release path readnext_state=None; maintainer adopted theprev_stateapproach and credited me as co-author.
NousResearch/hermes-agent (~216k★) — Nous Research's agent framework
- PR #64771 (merged): cost-safe model routing — a bare-provider
/modelswitch was silently routing to the priciest flagship model (a billing footgun that once escalated to a premium model and billed 863 requests before the user noticed); routed it through the cost-safe default instead, with regression tests. Merged intomain. - PR #61835 (merged): fixed a crash on null web/backend configuration and added regression tests; merged into
main.
BerriAI/litellm (50k★) — LLM gateway/proxy unifying 100+ providers
- PR #29707 (merged): Diagnosed a Vertex AI context-caching 404 on multi-region (eu/us) endpoints — the caching path hardcoded the single-region host instead of the multi-region REP host the inference path already used — and contributed the merged parametrized regression suite locking the corrected host-resolution invariant. 49 green CI checks.
vllm-project/llm-compressor — vLLM's model-quantization toolkit
- PR #2797 (merged): Added IBM Granite (
GraniteForCausalLM) to the AWQ and SmoothQuant quantization mapping registries, with meta-device tests that instantiate the model skeleton and validate mapping regexes against the real module tree. Also unblocked the merge itself by resolving atest_utils.pyrebase conflict that the maintainers' automation couldn't.
vllm-project/production-stack — official Kubernetes deployment stack for vLLM
- PR #969 (merged): Router bug fix —
route_sleep_wakeup_requestconsumed only the router-internalidquery param and silently dropped the rest, soPOST /sleep?id=X&level=2degraded tolevel=1. Fixed by forwarding all non-idquery params to every upstream call. - PR #976 / PR #970 (merged): Cross-platform macOS support for the cluster install tooling —
uname-based OS/arch detection,sysctl -n hw.memsizefor Darwin memory sizing, Linux-only calls gated behind OS checks.
| Project | Description | Stack |
|---|---|---|
| RepoAgentBench | Open-source CLI that mines merged GitHub PRs into reproducible, contamination-free coding-agent benchmarks — public benchmarks overestimate agent capability by 20–50% through training-data contamination. Surfaced what leaderboards can't: the same model produced opposite outcomes on an identical task under two different agent harnesses. | Python, PyPI, GitHub API |
| SafetyCommander (Zapdos Labs × Antler hackathon) | Factory-safety agent where a VLM judges risk by reading the written safety policy and citing the controlling clause — risk is decided in exactly one auditable module; edit one line of policy and the verdict flips. 📝 Architecture write-up | Qwen3-VL, vLLM, YOLO, RAG |
| Archiagents | Live AI product for architectural design (2-person team): briefs + CAD/IFC in → design schemes, photorealistic renders, IFC4 BIM models out. Competed in an OpenAI hackathon. | Vercel AI SDK, gpt-image-1, Autodesk APS |
| LLM API Gateway | Production multi-provider LLM proxy (Claude/GPT/Gemini) with load balancing and key management — powers my downstream products. | Node.js, Express, VPS |
| SmartStudy Agent (Chrome Web Store) | Closed-loop learning agent with a 4-policy benchmark — honestly reported that a rule-based heuristic (+35%) beat Q-learning (+18%) short-horizon, so the heuristic shipped. | Python, Claude API, Chrome MV3 |
| NeuroUnfold | Physics-informed DL recovering 406 kHz LoRa chirps from 5.3× aliased BLE RSSI at 0.986 R² — a signal classical sampling theory says is unrecoverable. | PyTorch, NumPy |
| 77 kHz BLE Firmware | Custom Zephyr RTOS firmware on nRF54L15 — 3× the highest published sampling rate, <0.01% drop. | C, Zephyr RTOS, DMA |
| Dual-Stream Gesture Transformer | Real-time gesture recognition at 557 FPS (1.79 ms), 88.2% accuracy from 35 labeled samples via sim-to-real. | PyTorch, MediaPipe |
- First author — When Memory Updates but Behavior Does Not: Repairing Implicit Stale Dependencies in Personalized Agent Responses · arXiv:2608.01619
- Robotic manipulation RL — sim-to-real on Franka & xArm, contact-rich policies in Isaac Lab
- Peer Reviewer — AgentSkills Workshop @ ACM CAIS 2026 · IEEE Wireless Communications Letters
- Member — AAAI · IEEE-HKN |
- 2 Chinese patents · Provincial 2nd Prize, China Undergraduate Mathematical Contest in Modeling
