welldun·Writing
Notes from building infrastructure for AI agents.
We build open-source infrastructure for AI agents that do real work, and write up how it works: the bets, the gotchas, and what we measured.
Series
one per projectSeries
Drafting Table
Building a multi-tenant platform where every customer designs their own database — the entity platform.
Series
Offhook
Building an open-source, self-hostable API for AI phone calls: STT, an LLM and TTS over a real phone line.
Series
SoonEmail Infra
A multi-tenant email-sending API — BYO SendGrid/Resend, envelope encryption, a Postgres work queue, signed webhooks.
Latest writing
8 posts- Offhook Building an open-source API for AI phone calls Why we built Offhook, what makes an AI phone call hard, and what this series will cover.
- Offhook Latency is the product Where the time goes in one spoken turn, why the model's first token decides how a call feels, and the fixed cost we hadn't counted.
- Offhook The phone part is the hard part What actually happens when an API places a phone call, and five ways two carriers doing the same job still managed to surprise us.
- Drafting Table Building a platform where customers design their own database Why we're building a 'define-your-own-data' platform — and what this series will cover.
- Drafting Table One engine, many schemas The core bet: a registry as source of truth, JSON for storage, and generated SQL views on top.
- Drafting Table Keeping tenants apart with Row-Level Security How Postgres makes data isolation a property of the database — and the view gotcha that nearly leaked everything.
- Drafting Table Unique keys and foreign keys… on JSON How to enforce real constraints and relationships when the data lives in a JSON blob and the schema was invented by a user.
- Drafting Table What a load test taught us A bottleneck we couldn't see, a 400ms query that became 0.3ms, and why 'a lot of data' isn't the same as 'slow.'