WeldCost: AI-Powered Cost Estimation Platform
The backend is FastAPI with JWT auth over PostgreSQL. The biggest win was caching. Redis in front of the hot read paths took p99 latency from 1.2s to 360ms. That is the difference between an engineer waiting on the tool and an engineer just using it.
The estimation runs as LangGraph workflows with RAG over the welding standards, deployed on Fly Machines. It automates about 90% of the welding calculations that used to be done by hand, which gives the team back 12+ hours a week. The design goal was trust, not novelty. It routes on intent, runs its tools deterministically, and keeps state explicit, so the same inputs produce the same estimate and every output can be traced.
Long sessions get expensive because the context keeps growing. I kept input size flat with rolling context windows, automatic summarization, and state-first prompting, which cut inference cost per request by about 20%. To stop the model changing behavior between deployments, I built a Pytest and LangSmith suite across 60+ cases with human-in-the-loop interrupts and a regression monitor. That cut workflow regressions by 75% and made weekly releases routine.
