Notes
The technical version.
Four essays on problems I have actually run: long-context memory on a small machine, why browser agents waste tokens, what a paid diagnostic finds, and the seven ways production AI fails. Every one includes the numbers and the parts that did not work.
Not a blog. No newsletter, no weekly cadence, no comments. Subscribe by RSS if you want the next one.
01The seven ways production AI failsA field guide to the failure modes I look for in a diagnostic — and the first fix for each one.Seven repeatable failure modes: the half-working vendor tool, the manual document workflow, classification at scale, the agent that stops on record 41, the runaway bill, the expensive wrong answer, and the unowned roadmap.November 5, 2026
12 min02What a paid AI diagnostic actually findsThe document is not a strategy deck. It is the arithmetic, the failure modes, and the three decisions that matter — before you spend the build budget.What the two-week AI Production Diagnostic contains, what it typically finds, and why it comes before a build. Includes the cost model, the eval gap, and the go/no-go.November 2, 2026
8 min033–10× fewer tokens: index the page, don't describe itA browser agent that sends element references instead of DOM, screenshots, or coordinates — and why the token bill is an architecture decision.How Tappi Browser reduces browser-agent token usage by indexing interactive elements, returning bounded observations, and keeping the action surface typed. Includes the token arithmetic and the failure modes the design removes.October 28, 2026
9 min04Long-context memory is a cache problemWhat a 1.1M-token session taught me about KV stores, attention probes, and the limits of a bigger context window.A technical note on Kamaloki: writing a session's KV cache to disk, probing it for relevant spans, and splicing them back into a small model's context. Includes the measured results and the parts that did not work.October 20, 2026
11 min
12 min02What a paid AI diagnostic actually findsThe document is not a strategy deck. It is the arithmetic, the failure modes, and the three decisions that matter — before you spend the build budget.What the two-week AI Production Diagnostic contains, what it typically finds, and why it comes before a build. Includes the cost model, the eval gap, and the go/no-go.November 2, 2026
8 min033–10× fewer tokens: index the page, don't describe itA browser agent that sends element references instead of DOM, screenshots, or coordinates — and why the token bill is an architecture decision.How Tappi Browser reduces browser-agent token usage by indexing interactive elements, returning bounded observations, and keeping the action surface typed. Includes the token arithmetic and the failure modes the design removes.October 28, 2026
9 min04Long-context memory is a cache problemWhat a 1.1M-token session taught me about KV stores, attention probes, and the limits of a bigger context window.A technical note on Kamaloki: writing a session's KV cache to disk, probing it for relevant spans, and splicing them back into a small model's context. Includes the measured results and the parts that did not work.October 20, 2026
11 min
Why publish these
Proof of thinking, not a content strategy.
A high-ticket buyer does not need another weekly blog. They need to see how you reason about the exact problem they have. These essays exist for that check — and as the material I would use in a client conversation.
Have the same problem?
If your system fails on long context, cost, or reliability, start with the 90-minute session — the fee is credited toward the diagnostic.
Book the session — $1,500