About
The homepage is the short version of what I ship. This page is the slower version: why I care about dependable AI systems, and why my work tends to orbit the parts between a promising model and a tool people can trust.
I am Peerapon Wechsuwanmanee, currently Senior AI Engineer / Machine Learning Engineer at REWE Group. I work on GenAI, agentic systems, semantic search, MLOps, and the operational details that make those systems useful beyond a demo.
Robotics beginning
Long before production GenAI, I was on an autonomous robot-soccer team, owning mechanical design, manufacturing decisions, and maintenance. That early lesson stayed with me: the unglamorous reliability work is what lets the rest of the team move fast.
2011–2021
At RWTH Aachen, and later as a research scientist, I modelled damage and formability in advanced steels. Simulations fail loudly when your assumptions are wrong; that is good training for machine learning.
Sep 2021 – Jun 2022
At Botnoi, I led a speech-to-text team working on Thai ASR. It was a reminder that AI is not abstract infrastructure: for many people, it is whether technology understands their language at all.
Jul 2022 – Mar 2024
At Data Reply, I moved across MLOps accelerators, computer vision, predictive maintenance, and Spark pipelines. The common pattern was not the model family; it was turning uncertain requirements into something a business could use.
Sep 2024 – Present
At REWE Group, I work on GenAI products, multi-agent systems, semantic search, and internal AI adoption. The challenge is no longer only whether a model can answer; it is whether the whole system can be trusted by employees, operators, and other engineers.
What I am thinking about now
Working principles
Monitoring, evals, deployment paths, migrations, and documentation are not support work around AI. They are the reason an AI feature survives contact with real users.
I like designs where the data flow, failure modes, and human override points are visible. If only the builder understands why it worked, it is not ready yet.
Adoption depends on trust, team habits, incentives, and review processes. A technically correct agent can still fail if people cannot safely fit it into their work.
Teaching
Explaining AI to learners, career switchers, and non-specialists keeps me honest. If I cannot explain the system clearly, I probably do not understand it well enough yet.
Lead Instructor · May 2024 – Jun 2024
Lead Instructor · Nov 2023 – Jan 2024
Instructor · Aug 2020 – Sep 2020
Outside the job title
Outside work, I like badminton, Dota 2, Japanese music and TV series, and finding restaurants worth returning to.