leduy.py_

research · long-form notes

Notes from the production floor

Long-form writing on AI engineering, production systems, and the honest gap between a benchmark and a bank customer counting seconds.

1 / 6

Every 30 seconds · pauses while you interact.

Source fragments passing through four verification gates before publication or review.
8 min read

Agentic AI changed what I have to prove before a question ships

#agentic-ai#evaluation#production

At Growtrics, the hard part was not making an agent produce more Q&A. It was deciding which answers deserved to exist, at what coverage and cost.

read note→
OCR output degrading on a difficult bank document — the moment a fine-tune isn't enough.
5 min read

Why we re-trained the OCR backbone instead of fine-tuning

#ocr#ml#production

Fine-tuning looked faster on paper. The mistakes kept returning in exactly the places that mattered.

read note→
A schema-aware Text-to-SQL agent: real table on the left, verified SELECT on the right.
6 min read

Shipping a Text-to-SQL agent that doesn't hallucinate column names

#agents#llm#sql

The first version sounded brilliant right up until it invented a field that did not exist. Confidence is cheap when the schema is not watching.

read note→
Saigon at dusk, Tokyo at night — the path of an AI engineering career between the two.
6 min read

From SmartPay to GMO: 18 months of AI engineering in Vietnam

#career#vietnam#ai-engineering

Eighteen months is not enough time to become wise. It is enough time to lose a few illusions about speed, certainty, and what good engineering feels like up close.

read note→
Latency bars dropping ~40% across the long tail — the work that gives the product better manners.
5 min read

What 40% latency looks like when a bank is your customer

#performance#production#trust

Latency sounds abstract until someone on the customer call starts counting the seconds out loud. Then it becomes a feeling in your chest.

read note→
Drag to move · tap to chat · double-click for terminal