2026 · Open source contribution
Diagnosed macOS kernel panics (completeMemory() prepare count underflow in IOGPUMemory.cpp) when serving a 27B model at 64k context on 24 GB unified memory. Traced the root cause to step=256 KV cache reallocation churn — each mx.concatenate call during prefill is a new wired-memory allocation from Metal. Patched mlx-lm's KVCache and QuantizedKVCache to preallocate the full N-token buffer in a single mx.zeros + mx.eval call at creation time, then write in-place. Zero reallocations. Stable 65,536-token serving across a 10-turn stress test with zero crashes.
2026
4-stage automated pipeline for systematic CV dataset discovery on Hugging Face Hub. Stage 1: metadata harvesting via HF Hub + Dataset Viewer APIs. Stage 2: extension bucketing and fuzzy deduplication (difflib.SequenceMatcher, similarity ≥0.90). Stage 3: VLM-based visual classification (Qwen2.5-VL/Qwen3-VL) with majority voting across N random samples per dataset, plus heuristic annotation detection. Stage 4: zero-download Parquet schema inspection and disk-safe controlled download. Screened 500k+ items in under 5 hours. Fully resumable with incremental CSV appending.
2026
CLI tool that benchmarks a YOLO model across every export format available on your hardware — PyTorch/CUDA, ONNX, CoreML, TensorRT — and produces real FPS, latency, and accuracy-delta numbers. Motivated by daily INT8/TensorRT profiling work: every engineer answers "should I use ONNX or TensorRT?" by hand, inconsistently, and throws the results away. exportrace aims to be the Cinebench of CV model deployment — one command, reproducible benchmark report you can drop into a README. Waitlist open.
Sep 2024 – Dec 2024
Web platform that pools compute/RAM across multiple low-end machines into one usable resource for 3D rendering, game dev, and AI inferencing. Engineered secure sandboxed code execution using Docker with CPU/memory limits on AWS ECS Fargate. Implemented PostgreSQL connection pooling (PgBouncer) reducing connection overhead by 70%. Built scalable backend supporting 1000+ concurrent users with strict security isolation.
Aug 2024 – Oct 2024 · Top 10 National Finalist, Smart India Hackathon 2024 (50,000+ teams)
On-device pose estimation (Google ML Kit) on Android for low-latency, privacy-preserving fitness evaluation. Firebase-backed pipeline for nationwide athlete profiling and leaderboards. Built during Smart India Hackathon 2024, reached national top 10 out of 50,000+ competing teams.