SKC Engineering Ltd
AI/ML Engineer
Overview
Estimating a weld job is a conversation — you gather parameters one piece at a time — and a naive LLM either forgets what you told it or drifts between runs. WeldCost is built to be dependable rather than clever: it turns that back-and-forth into a consistent, auditable estimate.
What happened
WeldCost - AI-Powered Cost Estimation Platform
I built it on GPT-4o mini and LangGraph, but the design goal was trust, not novelty. It routes on intent, runs its tools deterministically, and keeps its state explicit — so the same inputs produce the same estimate, and every output can be traced back to how it was reached.
Long estimating 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 quietly changing behavior between deployments, I built 50+ LangSmith eval cases with human-in-the-loop interrupts and a regression monitor that flags drift before it ships.
On the infrastructure side it runs on Fly.io behind Docker multi-stage builds and GitHub Actions CI/CD, with JWT access control and idempotent APIs backing a persistent multi-user workspace. The core authenticated endpoints answer in under 300ms.