All work
Computer Vision · DP World · 2025
On-Prem VLMs for Fire & Smoke
Deployed and evaluated Llama 90B vision-language models on-premise to detect fire and smoke in scenes where classic detectors struggle.
01 The problem
Conventional detectors miss fire and smoke in difficult scenes — haze, glare, night, clutter — and footage could not leave the premises.
02 What I did
- Deployed Llama 90B VLMs on on-premise GPU infrastructure.
- Built prompts and test sets covering the hardest real-world scenes.
- Benchmarked VLM output against conventional detection approaches.
03 Impact
- Better coverage of hard edge cases
- Fully on-prem — no footage leaves the site
- A reusable blueprint for VLM-based detection
Client work — details generalised to respect confidentiality. Happy to discuss specifics in an interview.