I build AI systems that can see, reason, act, and be audited.
Multimodal AI · Computer Vision · Agents & RAG · Full-stack AI Engineering
I care about the part after a model demo works: turning it into a system that is observable, reviewable, and useful in the real world.
|
SEE VLM, detection, OCR, image understanding |
REASON Agents, RAG, specialist review, orchestration |
BUILD Python APIs, Vue interfaces, model services |
SHIP Docker, Nginx, inference tracing, deployment |
|
Evidence-first multimodal fire-safety inspection. VLM and YOLO collaborate through specialist review, deterministic harnesses, traceable history, and human feedback.
|
A visual inspection system that combines YOLOv8 with language-model reasoning for cinema behavior and compliance analysis.
|
|
An agent project exploring the ReAct loop: reasoning, tool use, observation, and iterative decision-making.
|
Experiments around retrieval workflows that let agents decide when to search, verify, and refine their answers.
|
Evidence before confidence. A convincing answer is not enough; important conclusions should remain connected to inspectable evidence.
- Models are components, not the whole product.
- Observability is part of intelligence, not an afterthought.
- A specialist should only modify the conclusion it owns.
- Good AI interfaces make uncertainty visible and correction easy.
- The last 20% is usually deployment, data quality, and failure handling.
- Making multimodal inspection systems more reliable through evidence isolation and targeted review.
- Connecting visual models, detectors, agents, and deterministic logic into maintainable products.
- Building evaluation loops from real failure cases instead of optimizing only for demos.
Build systems that do more than predict. Build systems that can explain what happened.