I specialize in small language models that run on everyday hardware: multilingual fine-tuning, distillation into sub-2B students, and quantized exports, plus agent systems with human approval built in.
Ranked for AI depth, completeness, and demo-ability. Training and eval numbers included; ask me for the runs.
Offline farm advisor for Nigerian smallholders: a multilingual 3.3B tiny LLM plus an 18-crop disease classifier, both running fully on-device.

Type a drug once and nearby pharmacies get AI voice calls checking stock and price, ranked cheapest-first. A human approves every step.

Offline preaching projector that listens to the sermon and displays cited verses full-screen, with 96.5% exact-match on references.

Privacy-first LLM app on the Snap Store: a distilled 1.7B reasoner running offline at 20 tok/s, plus BYOK access to Claude, Gemini, and GPT.

Fine-tuned YOLOv8-OBB for 3D Rubik's-cube tracking under dynamic lighting: 95% mAP on a self-curated set, active-learning relabel loop, ONNX export.

Coding assistant that lives in your terminal: agent loop + tool registry + validators, file patching, command execution.
Strongest where training meets shipping: data pipelines with quality gates, evals that catch regressions (bias audits, forgetting checks), and quantized exports that actually run on the target device.

Also hiring elsewhere? Frontend portfolio · Full-stack portfolio