As founding engineer at a US developer-tooling startup, I designed the knowledge-graph backbone for an automated software-analysis platform - turning raw context into structured graph slices that made a fleet of AI agents both cheaper to run and more accurate.
The platform had to analyze codebases across 20+ languages and answer deep, cross-file questions. Feeding raw context to LLM agents was expensive and noisy - token budgets ballooned and answers drifted whenever the relevant facts were scattered across files.
I designed a three-layer knowledge-graph pipeline in FalkorDB (Cypher, 149 entity types) and had each of 7 specialist agents query a tailored graph slice instead of raw context. I built 20+ Temporal workflows and 120+ activities for durable orchestration, with Kafka streaming, ClickHouse analytics, and full observability running on production GKE.
Graph-scoped retrieval cut the tokens each agent consumed and measurably lifted output quality - the model saw only the connected facts it needed. The platform runs in production with durable, observable workflows the team can reason about and extend.
I take on a small number of engagements like this. Let's talk about yours.
Book a call