For 16 months I was the only engineer at my own industrial software company. I designed it, built it, sold it, and listened when customers said no. Two of the projects below are dead. The write-ups explain why, because I learned more from killing them than from building them. The newest one, an applied AI (RAG) support assistant, is live.
01
Work
A support assistant that answers product questions only from public documentation, cites the page, and declines when the documents do not cover it.
SEP 2026 – OCT 2026 · SOLE ENGINEER · STATUS: LIVE · DEMO: bas.jasonkhaings.com · CODE: github.com/jkhaings/bas-assistant
PYTHON · FASTAPI · POSTGRESQL/PGVECTOR · LANGGRAPH · LITELLM · REACT · GRAFANA
A retrieval (RAG) assistant: it searches the documents first, and code checks every citation before an answer is shown. It passes 22 of 22 checks on a fixed question set, and the write-up says what that number does and does not prove. It is an independent demo built on Delta Controls' public documentation, with no affiliation.
Read the case study
A vibration analysis web app for industrial machinery, sold to vibration analysts as a tool.
MAY 2025 – SEP 2026 · SOLE ENGINEER · STATUS: SHUT DOWN · LIVE: phasedraft.com · CODE: github.com/jkhaings/vib-agent
PYTHON · NUMPY/SCIPY · FASTAPI · POSTGRESQL · WEASYPRINT · STRIPE
A deterministic Python core computed every finding, and a language model drafted only the narrative text, checked against those findings before a report shipped. Nobody who tried it used it a second time, and the write-up explains why.
Read the case study
A managed predictive-maintenance service for BC plants: wireless sensors on their machines, monitoring and alerts handled by me.
APR 2026 – JUL 2026 · SOLE ENGINEER · STATUS: SHUT DOWN · COMPANY: khaingsinternational.com
NCD GEN4 SENSORS · NODE-RED/FLOWFUSE · INFLUXDB · FASTAPI · ISO 20816-3
Five analytics layers ran over wireless sensor data, deployed as a pilot at a dairy plant. Small plants did not care enough to pay and large plants stayed with incumbent vendors, which made the credibility gap the real product problem.
Read the case study