Wrote a fused GroupNorm + SiLU CUDA kernel for diffusion U-Net backbones, cutting HBM traffic from 5 to 3 array-sized passes via block-per-group warp-shuffle reductions, fp32 accumulation, and vectorized 128-bit loads verified in SASS.
Benchmarked against PyTorch eager, torch.compile, and a Triton port of the same algorithm: 2.3–3.5× over eager and 1.1–5.0× over torch.compile at 48% of peak HBM bandwidth.
Predicted 1.67× from a memory-traffic count and measured 1.3–3.5× instead. The spread falls exactly on L2 capacity — below it the intermediate never reaches HBM so fusion saves little, above it every implementation converges on the memory system’s limit and the code stops mattering.
Extended aerial vision-language navigation with battery awareness on the AerialVLN benchmark — 32.4 GB of AirSim/Unreal flights across 25 city-scale environments and 870+ object types — annotating every timestep with a normalized fuel fraction so the policy could condition on remaining charge.
Compared Seq2Seq and Cross-Modal Alignment (CMA) policies trained with teacher forcing and DAgger. CMA cut navigation error 21.4% on unseen environments, 218.9 m to 172.1 m, and lifted success-weighted nDTW from 0.7 to 1.1.
Built a FuelGateModule with an auxiliary MSE loss to trade exploration against goal-seeking — then found it barely moved the needle. Imitation learning already encodes efficient trajectories, and a scalar fuel signal is drowned out by the visual and linguistic features feeding the policy.
Built a deployed Django web app where players collect, trade and quiz on mythical creatures from world cultures — eight apps covering authentication, collections, a luck-based draw, a marketplace, a myth map, quizzes and notifications.
Modelled ownership as a many-to-many relation through a join table so the same creature can belong to many players with per-player state, and built a two-sided Deal model that makes peer-to-peer trades atomic rather than two independent transfers.
Wired creature rarity into gameplay: draws are region-scoped and probability-weighted, and creatures grant quiz power-ups — time boost, eliminate answer, replace question, jump ahead, skip after streak — so collecting feeds directly back into competing.
A feature store aimed at training-serving skew — the failure where a model sees one definition of a feature in training and a subtly different one in production, and quietly degrades.
The design: define each feature once as a versioned SQL template, freeze a feature-set version, backfill it offline to build training matrices, materialize the latest values online for inference, then run consistency checks that report per-feature mismatch rates between the two stores.
FastAPI service over a Postgres schema managed with Alembic migrations, with a Next.js frontend wired to the API.
Built a hybrid retrieval pipeline over a 100-paper arXiv cs.LG corpus: PDFs parsed into section-structured JSON, chunked two ways (fixed token windows and a semantic cascade falling back section → paragraph → fixed), and indexed as 5,068 chunks in Qdrant.
Combined dense embeddings (all-MiniLM-L6-v2 via fastembed) with sparse BM25, min-max normalizing each within the retrieval pool before an alpha-weighted merge — so the two scoring scales stay comparable instead of one silently dominating.
Wrapped the LLM client with exponential-backoff retries, Pydantic-validated structured outputs behind a self-correcting retry loop, and JSONL tracing. 22 tests cover retry logic, schema validation, hybrid scoring and tokenizer edge cases.
Next: agent orchestration over these tools, then an eval harness measuring precision@k, latency and answer consistency.