portfolio · linkedin · email

CS @ UCF (Burnett Honors College), undergrad research fine-tuning LLMs. I like the layer under the model — training stacks, GPU kernels, tokenizers, data pipelines — and I'd rather write these systems than read about them.

_{C++ CUDA Metal Python PyTorch JAX TypeScript}

Pretraining Shohin — a 135M-parameter language model built from scratch (custom tokenizer → custom data pipeline → 8×H100 training run), targeting state-of-the-art reasoning at the ≤150M scale. Release and writeup soon.

Also on the bench: gods-eye — real-time monocular 3D reconstruction in Rust: one RGB camera in, a live adaptive-LOD triangle mesh out. Design stage.

psi — a small language model and the training stack it runs on, written from scratch in C++: own reverse-mode autograd (gradient-checked to ~1e-12), own Metal and CUDA kernels, zero ML frameworks anywhere in the stack.

• the 354K-param model matches the original TinyStories-1M on grammar / coherence / consistency / plot — at a third of the size
• hand-written Metal matmul: 883 GFLOP/s at 2048³ on an M1 — 34% of peak, 6× the naive kernel, bit-exact against the CPU path
• ternary-weight matmul ({−1, 0, +1}): 16× smaller than fp32, running at full fp32-matmul speed

sai — voice-native agentic OS co-pilot built on Amazon Nova: it hears a command, sees the screen (pure vision — no DOM, no selectors), and acts through a Plan → Act → Verify agent loop. "Hey Sai, answer this LeetCode problem" → reads it off the display, writes the solution in the editor, clicks Submit. 🏆 Best Agentic AI @ Amazon Nova AI Hackathon (13,000+ participants).

Project Sentinel — real-time voice anti-fraud guardian: streams live call audio over full-duplex WebSockets, transcribes with Deepgram, runs sub-second threat analysis on Cerebras (Llama 3.3-70B), and alerts mid-call — before the scam lands. 🏆 Winner @ The AI Championship.

Crisis Topography — voice-commanded 3D geospatial crisis-intelligence platform: agentic orchestration over Databricks analytics and a self-hosted Actian vector store (18k+ embeddings, sub-100ms retrieval). 🥉 3rd overall @ Hacklytics 2026 (Georgia Tech, 1,200+ builders).

vectorized-2048 — an RL environment rewritten as stateless, branchless matrix ops in JAX: 4,096 games simulated in parallel, 200K+ training steps/sec on an M1 Air (2M+ inference), simulation and DQN training end-to-end on-GPU with zero host transfer.

Merged PRs in TensorFlow core — a segfault guard in TensorListSetItem and a float32 erfinv precision fix near ±1 — plus Unsloth, LiteLLM, torchtitan, and wit-bindgen.

→ every merged PR, live from GitHub search — no curation, just the query.