ML systems engineer · GPU performance · LLM inference · AI agents whose work can be checked
Research Scientist at the Research Foundation for SUNY · San Francisco Bay Area
I make large models train and serve faster, and I build AI agents whose work can be checked.
- 🔬 Now: Research Scientist at the Research Foundation for SUNY. I profile and tune LLM fine-tuning on NVIDIA B200 (Nsight, Triton and CUDA kernels, FSDP, QLoRA) and build deep-learning models for genomics.
- 🏗️ Before: ML engineer at A5G Networks (5G traffic forecasting, Databricks and Delta Lake pipelines) and at Cloudphysician (real-time ICU alerting on Kafka, PySpark and AWS); data engineer at InxiteOut.
- 🧪 On weekends: hackathon builds on agent security, computer use, video agents, rare-disease diagnosis and cryo-EM.
- 📄 7 peer-reviewed papers and 3 patents, listed on Google Scholar.
|
Zuzu: call a number, answer plain questions in your own language, and six agents on BAND fill in a USCIS form, with every value taken from a deterministic tool. Built with Gar Walsh.
🏆 Best Use of BAND and Best Use of ElevenLabs, Prompt Driven Hackathon 2026 |
|
|
CryoClear: a real-time cryo-EM junk-triage copilot. It picks particles, flags junk as micrographs stream in, and learns from a scientist's corrections. I built the picking engine, the junk classifier and the active-learning UI.
QBI Hackathon 2026, UCSF · open source (MIT) |
|
|
Diageno: a rare-disease diagnostic recommendation prototype. From phenotype and genetic data it ranks a differential diagnosis, shows the evidence and uncertainty, and suggests the most useful next question.
Harvard Rare Disease Hackathon 2026 · team of four |
Dao-AILab/flash-attention#3102:
mha_fwd_kvcache now checks the length of cache_batch_idx, so a short index raises an error instead of reading past its end (fixes #3097). Merged, along with the same check for FA3 in #3153.
smg-project/smg#5353: Harmony chat prompts now put a leading system message in the developer message and leave out an empty one, as vLLM does (#2830). Merged. #5354, in review, makes
least_load count its own picks on workers that send no load report (fixes #4602).
Slicer/Slicer#9513: removes redundant variables in the GrowCut and Terminologies logic (fixes #9462 and #9463). Merged. #9515, in review, adds a test for the slice view N-layer API (#8294).
LegalQuants/lq-ai#783: constrains the Teams install-only
id_token flow (D-03). Merged.
linkedin/Liger-Kernel: kernel fixes checked against PyTorch references. Merged: the DyT forward launch above 65535 rows (#1607) and the column stride in the fused KL and JSD gradients (#1613), both from a B200 sweep. In review: GroupNorm backward kept in fp32 so fp16 gradients stop turning into NaN (#1608), and label smoothing in vocab-parallel cross entropy, which returned inf when a logit was far below the row max (#1619).
llm-d/llm-d-inference-sim: three fixes from a parity check of the simulator against real vLLM: the time-per-output-token metric no longer skips single-token outputs (#891),
request_params_max_tokens no longer skips requests without max_tokens and records the value each request ran with (#894), and a 400 from the render service now reaches the client as a 400, not a 500 (#898). A fourth fix stops the vLLM metrics adapter's unit tests from starting the adapter twice, which made them race on its request counters (#934). All merged.
llm-d/llm-d-benchmark#2149: adds demand p95 and an optional demand series to the sizing report (#2134). Merged.
nf-core/modules#13483:
utils_nfcore_pipeline keeps trailing zeros in software versions, so samtools: 1.20 no longer turns into 1.2. Merged.
Nimbleway/webrouter: two bugs I found in WebRouter, which connects AI agents to the web, each fixed with regression tests: search through exa-mcp and you-mcp no longer fails when a result contains U+2028, since event streams now split lines only at CR and LF (#353), and
auto extract keeps pages in the order asked for after it retries a failed URL (#354). Merged.
ai-dynamo/dynamo: five bugs I found in Dynamo's KV routing, runtime and request handling, each fixed in its own PR with a test that fails on main: binary extra keys in KV events (#18473), repeated block events in the standalone indexer (#18476), the QUIC response plane during a graceful drain (#18477), a 503 instead of a 500 when no worker is eligible (#18479), and stopping the worker once every choice has finished (#18481). In review.
triton-lang/triton: bugs on Blackwell from my B200 test sweeps. Fixes in review: wrong
tl.dot results for a transposed operand above 256 (#12672) and 8-bit dots with 8-element rows that fail to compile (#12674). Also reported: #12675 and #12678. Two earlier sweep bugs, #12606 and #12607, were fixed upstream.
NVIDIA/TransformerEngine#3790: adds
export_components() to MXFP8 tensors, returning the FP8 values and their compact scales (#3644). In review.
BristolSTA/uobtheatre-api#1044: stores emails in lower case and matches them ignoring case (fixes #834). In review.
NVIDIA-NeMo/Curator#2285: showed that the fuzzy-dedup tuning guidance for
num_bands and minhashes_per_band was inverted under MinHash-LSH banding. The maintainers fixed the docs in #2287.
spack/spack-packages#7083: makes charliecloud 0.44 and later require squashfuse 0.4.0 or newer (fixes #5400). Merged.
promptdriven/pdd#2341: context limits are handled as a hard cutoff and provider prompt caching is never used; with an RFC on serving assumptions in #2342.
COMPAS-Lab/cloud-manager: 15 merged pull requests to a research lab's cloud console, a fork of Linode Cloud Manager (TypeScript, React).
- high-level-design: system design notes and case studies, including ML and AI-infrastructure designs with capacity estimates
- low-level-design: 48 runnable object-oriented designs, from a parking lot to an inference server, a KV-cache pool and a GPU memory allocator
- leetcode: 935 solutions in Python, C++, Java, C, SQL and bash
- Study path: my B.Tech coursework repos, arranged as a curriculum
![]() Dune Buggy Sandbox Bullet vehicle physics on procedural dunes, with ramps and loop rings |
![]() Orbital Defense Top-down arcade shooter with pickups, a shield and an animated GLSL starfield |
![]() Beam Bender Grid puzzle: rotate the mirrors to bend the beam onto every target |
![]() Still Life Painter Orbit a 3D still life and paint what you see on the canvas |
qgis-geospatial-projects: six QGIS and Python analyses (urban heat island, flood risk, crop health, healthcare access, forest change and solar siting). Each has QGIS layer styles, a Python pipeline and a Streamlit app, and the maps are live on the project site.
![]() Urban heat island |
![]() Flood risk |
![]() Crop health |
![]() Forest change |
Animated emoji: Microsoft Fluent Emoji (MIT).

















