Hey — I'm a third-year CS student who spends most of his time trying to make small models punch above their weight. Lately that means RL post-training, inference-time memory, and figuring out what's actually possible on a laptop GPU.

I like problems where the constraint is the interesting part.

Qwen3-0.6B Reasoning Pipeline  ·  active

Training a 0.6B model to reason using a 4-stage GRPO pipeline — SFT coldstart, RL on math, mode fusion, then generalization. At inference I attach a Hopfield episodic memory bank (~20MB) that retrieves similar past problems as context. The bet is that a sub-1B model with the right inference-time setup can match 7B+ on reasoning benchmarks.

PowerBench-Consumer  ·  complete

Benchmarked LLM inference and GRPO training on an RTX 2050 (4GB VRAM) — extending DREAM:Lab's Jetson Orin AGX methodology to hardware most people actually own. The interesting finding: INT8 is 3× slower than FP16 on consumer GPUs, the opposite of what happens on the Jetson. Quantization benefits don't travel across hardware architectures.

FP16  →  2,407ms  ·  13.29 tok/s  ·  PPL 14.80
INT8  →  10,056ms ·  3.18  tok/s  ·  PPL 19.46
INT4  →  3,965ms  ·  8.07  tok/s  ·  PPL 15.36

Neural Global Illumination Engine  ·  complete

Reframed Global Illumination as a regression task — an MLP learns to predict radiance instead of ray-tracing it. Brought per-frame compute from 26ms down to 11ms with a +4.1dB PSNR improvement. Runs at 60+ FPS with dynamic lighting and moving occluders.

Deepfake Detector  ·  complete

Multimodal detection pipeline across video and audio. 85%+ accuracy, processes 10K+ frames and audio samples per batch in under 3 seconds.

{
  "research":  ["PyTorch", "GRPO", "QLoRA", "TRL", "HuggingFace Transformers"],
  "cv / edge": ["OpenCV", "TFLite", "YOLO", "Taichi CUDA"],
  "backend":   ["FastAPI", "Node.js", "Express", "MongoDB", "PostgreSQL"],
  "cloud":     ["AWS Lambda", "Bedrock", "S3", "DynamoDB", "Step Functions"],
  "languages": ["Python", "C", "C++", "JavaScript", "Go"],
}

Qualcomm Edge AI Hackathon 2025 — Finalist  ·  top teams from 2000+ participants
DevsHouse '25 — MongoDB Track Winner  ·  4th overall