← All work

AI summarization

Long-form proceedings, distilled

Hours of recorded sessions become a four-minute letter — every claim time-linked to the tape.

The system

An ingestion pipeline watches long-form recorded sessions nobody has time to watch, extracts what was actually decided, and threads related items into narratives that span months.

Every extracted claim carries a timestamped citation into the source video. Receipts, not summaries — a reader can check any sentence against the tape in one click.

The hard parts

Auditable summarization

The citation-first design makes model output verifiable. Trust comes from the link, not from the prose.

Threading across months

Items are resolved into long-running threads across sessions, so a decision's entire history reads as one timeline around a today-line.

Six days to a working system

Ingestion, extraction, threading, a reader, and the render-to-send email rail — zero to working in six days.

In numbers

90
sessions ingested
1,575
claims extracted
162
long-running threads
6
days to build

Python · caption harvesting · Next.js · email rail

Happy to walk through this one properly — what it does, how it's put together, and what it would take to build something like it for you.

Vague is fine.

What kind of help? optional — pick any