because
Scope, data, and a success metric — defined before any code.
Agents, retrieval, and tools — wired into one pipeline.
Evals, logs, production — measured, not demoed.
I build and ship production AI workflows — fine-tuning language models, wiring agent systems with tool use, and deploying retrieval-backed pipelines with PyTorch, LangChain, and FastAPI.
My work runs the full loop: framing the problem, integrating the APIs, evaluating the outputs, and getting it into production — from multi-step agent automation to reproducible image-generation workflows.
Every system below shipped with a number attached. Not demos — production workflows with before/after deltas, measured across real team usage.
→Four deployments, four deltascut in manual support effort via RAG agent pipelines
faster design iteration with LoRA + ComfyUI workflows
less training time via mixed-precision LoRA pipeline
faster designer onboarding across 4 shipped workflows
Four production systems — every one shipped with a number attached. Not demos: indexed codebases, agent memory, and diffusion pipelines running for real teams.
A context intelligence layer for AI coding agents — full structural awareness before a single file is touched.
linear-canal.comMulti-agent context management with layered memory — hot store, semantic index, durable persistence.
End-to-end domain-specific image generation, deployed as a concurrent inference service.
Turns agent sessions into durable beliefs tied to their source, time, and history — memory you can audit.
final year — in progress
Coursework weighted toward applied machine learning — from systems fundamentals to deep learning and computer vision.