production· client work
Two-stage vision-language inspection pipeline
A two-stage VLM + LLM pipeline that inspects and reviews submissions for a national tourism platform, served on 4× A100 80GB.
4× A100 80GB
Problem
A national tourism platform receives a high volume of image and text submissions that need consistent, explainable review. Manual review was slow and uneven.
Approach
- Stage one, perception. Qwen2.5-VL reads each image together with its context and produces structured observations.
- Stage two, judgment. Llama 4 Scout reasons over those observations against the platform’s rules and returns a decision with a short rationale.
- Serving. Both models are self-hosted on 4× A100 80GB with batched, high-throughput inference, so data stays in the client’s environment.
Results
- Consistent, rule-based decisions with a human-readable reason for every item.
- Two stages keep each model’s job narrow, which made prompts and evaluation simpler to iterate on.
Stack
Qwen2.5-VL, Llama 4 Scout, vLLM, Python, NVIDIA A100.
Client work. Details generalized and no code is published.