Factory AI Consulting

Factories are where AI stops being a demo and starts meeting paper, dust, scans, regulations, and actual operational bottlenecks.

The clean public version: I have been building AI systems for factory environments, with the strongest current case around knowledge-base agents. Think quality documents, HR regulations, amendments, scans, old PDFs, department-specific terms, and people who need answers with sources, not a chatbot that vibes in their direction.

The core pipeline is boring in the good way: OCR the corpus, normalize documents, build a searchable index, preserve document structure and amendment chains, give the agent domain-specific tools, wire messenger integrations, and evaluate it with questions that expose retrieval, factual completeness, citation quality, and answerability detection.

The massive part of the work was the harness and benchmark layer. This is where "chat with PDF" stops being a product category and becomes engineering: custom tool schemas, corpus-specific retrieval operations, model-judge evaluation, answerability tests, source-citation gates, and repeated harness versions until the model is forced to behave like a senior archivist instead of a confident storyteller.

The local-AI track became surprisingly deep. I tested whether on-premise models could get close enough to frontier systems for Russian document QA, which hardware actually makes sense, and how much quality comes from the harness rather than the raw model. Current internal result: local deployment is viable for this class of KB when the corpus, tools, prompt, and evaluation harness are built properly.

The practical target is simple and hard: a source-cited assistant that can survive real operational questions while running fully local on a Mac Mini / NVIDIA-class box when the client needs on-premise deployment. The frontier is not only the model. It is the whole system around it.

I am keeping the client identity and operational details private. What belongs here publicly is the pattern: factory AI should start with constrained, source-grounded systems that save expert time, generate reusable process knowledge, and build the trust needed for deeper local AI deployment.