Writing
Notes from the builds.
What worked, what we got wrong, and what we would do differently. Written by the people who did the work, which is why it is specific and occasionally unflattering.
Featured · Patterns
Cutting AI agent costs by putting the expensive model at design time
We helped a tax accountant in India get an AI agent into their filing workflow. It worked on day one and was on a cost trajectory that would not survive a filing season. The fix was not a better prompt. It was changing who does the thinking, and when. Seven mechanisms, one open-source toolkit, and the measured bill for each phase.
July 2026 · 14 min read · Open source
Measured cost, by phase
- Design time, frontier model
- ~$400
- First two returns, before the fix
- ~$50
- Per client, after the split
- ~$1
The first two figures are measured. The third is the design-basis projection, and stays labeled that way until a full client has run through.
Archive
Everything else we have published
-
The parts of programming that did not get automated
An agent verified every field and reported success. The feature was still broken. Five skills survived automation, and the thing that used to install them did not.
Patterns
August 2026 · 9 min read -
Building a capability finder that routes you to the right person
A capability finder for a multi-site research program. Every claim links to its source, and every answer ends at the person who owns the instrument.
Research centers
August 2026 · 9 min read -
Building a grant advisor that knows when not to generate
A proposal advisor grounded in the solicitation and the office’s own funded proposals. The good version generates less than you would expect.
Research centers
August 2026 · 8 min read -
Adding plain-English queries to a city data warehouse
A plain-English layer over a city department’s own data. The interface was the small part, and the decade of integration underneath is why it works.
Government
August 2026 · 8 min read -
Building a research scoring system that shows its work
A facilitated self-assessment across nine dimensions, built so a reviewer can walk backwards from any score to the evidence behind it.
Research centers
August 2026 · 7 min read
Next
In the queue
Drafted and being written up, roughly in this order. Every one is grounded in something we shipped, which is the only rule the queue has.
-
The annual report is the best AI project your center is not doing
Four weeks of coordinator time, every year, in every center in the country. Nobody writes about it because it is not glamorous.
Research centers
-
If a user cannot check the answer, the answer is a liability
Why we tune our systems to refuse more often than a consumer product would, and what that costs in satisfaction scores.
Patterns
-
Universities are not slow buyers. You are talking to the wrong person
A principal investigator with grant authority can decide in four weeks. Central IT cannot decide at all.
Research centers
-
What an audit trail for an AI system actually has to contain
Written after a review board asked us a question we could not answer, and what we changed so that we could.
Government
Reading this because you have the same problem?
Thirty minutes, no deck. You will leave the call knowing whether it is worth building and roughly what it would cost.