Production AI workflows on constrained hardware

Problem

Enterprise customers wanted AI assistance over their geospatial data: natural-language querying, automated hazard and risk analysis, detection workflows. Two things made this harder than a typical LLM feature: the answers had to be trustworthy enough to act on operationally, and for on-premise customers there was no cloud inference endpoint to call.

Constraints

• Non-deterministic model output was unacceptable for workflows customers make operational decisions from.
• Geospatial context routinely exceeded usable context windows.
• On-premise deployments meant constrained GPU hardware and no external inference API.

What I built

I built production AI assistant workflows using LangChain, LangGraph, and RAG pipelines over geospatial datasets, and designed deterministic inference orchestration so that the parts of a workflow requiring reliability are structured graph steps rather than free-form model discretion. Alongside it I designed context management workflows for constrained-context LLM systems, keeping retrieval relevant instead of merely large.

• RAG pipelines over geospatial datasets, with pgvector for retrieval.
• LangGraph orchestration with deterministic step boundaries around the parts that must not vary.
• Context management for constrained-context systems.
• A Docker-outside-Docker GPU proof of concept enabling AI workloads to run on constrained on-premise infrastructure.

Result

I led a cross-functional team of 6 engineers and a Data Science team to deliver the HRA and Smart Detect AI features end to end. Both run in production, including for customers with no cloud connectivity at all.