William Obino

I’m an AI/ML engineer focused on model training, inference tooling, evaluation, and multimodal accessibility pipelines. My work spans sign language ML, RAG systems, LLM tooling, imitation learning, reinforcement learning, and Linux-first systems integration.

At Ishara AI (May 2026 – present), I built and debugged a sign-to-text pipeline for continuous sign language understanding, including dataset preparation, CTC decoding, and evaluation consistency. I also ported local training and inference workflows to Intel Arc/XPU while preserving CUDA compatibility, using shared PyTorch decode utilities and centralized runtime device selection—then validated end-to-end learning with bounded local train/eval runs.

As an AI Engineer contractor (Jan 2026 – Apr 2026), I designed and built a production REST API for a Kenyan LMS using Bun, Hono, and TypeScript, including pgvector/PostgreSQL retrieval with citation-grounded tutoring over SSE. I shipped a fully Dockerised, CI-tested service with OpenAPI documentation, idempotency-keyed endpoints, and structured observability logging, and I delivered a containerised media subprocess service using Manim, ffmpeg, and full TeX Live. Outside client work, I build and contribute—founding AcetutAI and LauncGround and working on robotics/AI projects like ACT-style imitation learning and visuotactile perception, while also contributing upstream patches to open-source agent tooling and Intel/XPU-compatible LLM tooling.