Tee Lakkhananukun · Bangkok
Work

The work, and the decisions behind it

Most of this work comes from Chemical Express, where I built the connections between our business systems and later developed sales models and quotation AI. I've written about the problems we were trying to solve, the decisions I made and what it takes to keep the systems useful to the people relying on them.

For a first read, try why I made QT smaller or the date-picker request. If you'd like the technical detail, follow what happens when a customer changes a quotation request or a model through a release decision. The project pages below bring the full stories together.

QT AI AgentsHelping sales turn a complicated customer request into a quote, while keeping track of the details that need checking.
  1. Operating agents in production
    Following email replies, remembering progress and knowing when customer service needs to take over.
  2. Keeping the work moving
    Fargate workers, failed tool calls, stale queued work and an estimated operating budget.
  3. Releasing a change to the AI
    Checking product matches against past requests, learning from review feedback and planning the next tests.
  4. When a similar product is wrong
    Finding the right item in a large catalogue takes more than matching the name on the bottle.
  5. Making QT smaller so people would use it
    Pulling back the first release, introducing product search on its own and learning from the team's corrections.
Sales recommendation platformHelping a salesperson work out which customers are worth contacting, what they might need, and why now is a good time.
  1. The model improved. Should we release it?
    How I would check whether a better test result translates into a model we can trust in daily use.
  2. What the model knew at the time
    Following late-arriving orders from source records into training data and daily predictions.
  3. When the recommendations feel wrong
    Tracing a sales complaint through data quality, delayed outcomes and model behaviour.
  4. From predictions to decisions
    Why nine models do not need nine live services, and where business rules belong.
  5. The salesperson who ignored the list
    A story about what a salesperson knows that the recommendation system does not.
Operations platformWorking through the purchasing, delivery and billing details that sit between receiving an order and getting paid.
  1. Why is the order still stuck?
    The different things sales, finance and the warehouse need to know.
  2. A system a small team can operate
    Where state belongs, how changes reach the ERP, and what it takes to recover safely.
  3. It started with a date picker
    How a small request led me to the underlying data, shared holidays and the other workflows that would depend on them.
EnvSearchA public English and Thai regulation-search project, with cited answers, an evaluation set and tools for Claude Desktop.
VerdaA pool villa being renovated for short stays, and the questions I am working through before opening.