17 min read
The product was the easy part
What a SaaS needs before it can charge anyone: credit billing in Stripe, invoicing and sales tax, email unsubscribe law, terms and privacy, and an admin panel.
Writing
Notes on building and scaling technology: AI and agent infrastructure, the architecture that holds up under load, and how enterprises actually get their digital costs down.
8 min read
The job was never the code
Most of my code is now written by an agent, and my output went up rather than down. That is not a story about typing speed. It is about what the job always was underneath the typing.
5 min read
Metering tokens when the bill is the product
Billing per message is easy and wrong: one user sends a sentence, another uploads a report. Metering the tokens you actually spend is harder, and the hard parts are idempotency, allowance checks and what to do mid-conversation.
4 min read
Route cheap models to orchestration, not to the work
Most agent platforms pick one model and use it everywhere. Splitting model selection by role rather than by task is where the cost curve actually bends.
5 min read
Citations that survive the question
A RAG answer with a source name under it is not a citation. If a teacher cannot open the page and see the sentence, the system has not shown its work. Carrying page numbers through retrieval is most of the job.
7 min read
One Durable Object per agent
Why I gave every AI agent its own Durable Object with a SQLite task ledger, and what that bought in state, concurrency and debuggability.
6 min read
Sandboxing code an agent wrote for you
Once an agent writes and runs code, you are executing untrusted input on your infrastructure. The isolation question has known answers. The harder problems are quotas, cancellation and getting results out.
6 min read
Persist once, reference by id, hydrate on demand
An agent that reads your uploaded file perfectly and then forgets it on the next turn is not a memory bug. It is a context design where file content was only ever turn-transient. The fix also cuts token cost.
4 min read
Reasoning effort is a cost dial, not a quality setting
Every major provider now exposes a knob for how hard the model thinks. Left at its default it burns tokens on arithmetic. Set per role, it is one of the largest cost reductions available without changing models.
5 min read
Crawling a knowledge base without a crawler service
Rendering every page in a headless browser is the expensive way to import a website. Discovery through robots.txt and sitemaps costs almost nothing, and most pages never need a browser at all.
5 min read
Confidence scores are only useful if something acts on them
An LLM that returns a per-field confidence score has told you nothing until a threshold routes the low ones to a person. How to design the routing, calibrate the threshold, and build a review queue people can actually clear.
6 min read
Signing webhooks, and the SSRF hole nobody patches
Most webhook guides stop at HMAC signatures. The bigger risk is on your side: a customer-supplied delivery URL turns your sender into a proxy for your own private network. Here is the full path.
6 min read
An event-sourced ledger a regulator can replay
Regulated lending needs a financial history you can reconstruct, not one you infer from mutable rows. Event sourcing plus double-entry bookkeeping gives you that, provided you get the arithmetic and the correction path right.
6 min read
Alarm-driven task queues without a queue service
A Durable Object alarm gives you a scheduler, a retry policy and exactly-one-worker-per-task, using storage you already have. Here is how to build a task queue on it, and where it stops being the right tool.