Blog

Notes from the build

Short, first-person pieces on what actually works when you put AI into production. Evidence over hype, no predictions.

Rules Model Route
FeaturedLLM engineering6 min

Hybrid routing: cutting LLM cost ~90% without losing accuracy

Sending every query to a model is accurate and expensive. Sending everything to rules is cheap and brittle. Here's the routing pattern I used to get most of both.

Read the note 18 June 2026
Document AI5 min

Making model output trustworthy enough to show a customer

The gap between a working extraction demo and a system a business will put in front of a customer is almost entirely validation. Here's how I close it.

02 June 2026
Evaluation7 min

Evaluating LLM outputs with ROUGE, BLEU and BERTScore

You can't ship AI responsibly if you can't say whether it got better. A practical look at the metrics I used to evaluate model output across 20k+ samples.

21 May 2026
Shipping6 min

Shipping bydhome from zero: what 0→1 actually costs

Co-founding a live marketplace taught me that the AI is the fun 20%. The other 80% is what separates a demo from a product people use.

06 May 2026