The brief · revised 19 September 2026

Measured AI systems, built fast and kept honest.

Madhu builds AI systems for documents, video and inference, and measures them before believing them — accuracy, latency and cost, with the negative results kept. Behind it, twenty years in the United States; in front of it, a research practice run in the open. Every figure below opens the file it came from.

0.9679

Document reading accuracy

ANLS on 300 DocVQA questions — every point from engineering, not a bigger model

84%

Fields filled with no human

and those are 94.1% exact; the rest go to a person with the evidence

−87%

Cost per AI answer

13.9 s → 6.1 s, open 30B model replacing a frontier API, quality held

40,116

Database benchmark score

from 9,047 — 108 experiments run by an autonomous loop overnight

$0.01

Per day to run that database

a free-tier VM, durable state in two regional buckets

0.917

Video Q&A retrieval

recall@10; a 37-minute talk ingested in 52 s for 4 cents

$3.50

Total spend, extraction research

41 numbered findings, including the ones that failed

Systems built

5 running in production today, toward a public goal of 100

Three bets

AI that knows which answers to trust

A reader that is 95% right and equally sure about everything saves no one any work. The research sorts answers into accept, re-check and ask-a-human — with a box on the page for every value — and that sorting is the product.

The thread →The findings →

Inference at a fraction of the price

Measured end to end on real traffic shapes: when an open model on rented GPUs matches a frontier API, what it costs per answer, and what streaming, caching and a rule engine each buy. Seventeen notebook parts and a sixteen-post series.

The notebook →The research thread →

Infrastructure that tunes itself

A document database whose truth lives in a bucket, optimised by an agent that proposed, benchmarked and kept or discarded 108 changes while its author slept. This site runs on it, and this site improves itself the same way.

The paper →The loop, run by run →How this site is made →

Pace — measured, not claimed

Hours at the bench, last 30 days	128on 23 days · 59.0 h this week · the log
Experiments recorded	133108 by the database loop, 25 in public notebooks · the loop · notebooks
Research threads with numbers	7each with its sample size and spend · research
Written and published	18essays since 1 April 2025, 112 lessons in 3 courses, 1 paper · about 466k words · writing · courses
Knowledge that is tested, not just stored	1509recall cards drawn from the work, graded against the source · 39 numbered findings · recall
Not public yet	3private projects, including an unlaunched company — shared by link on request · ask