Distill — a weekly intelligence digest on any topic you name

Strategy & Knowledge • Independent product

Read one: aidevtube.com — the AI-development digest this engine grew out of, published weekly and free to read.

What it is

Name a topic. I hand-pick the sources that actually matter for it — YouTube channels, podcasts, newsletters, news feeds, expert accounts on X — and from then on you get one edition a week: every item summarized from what it really said, condensed into a themed overview of the week, with a link on every claim. It is the answer to a problem most teams solve badly: the field you need to follow produces more each week than anyone can read, and skimming headlines tells you what was published, not what happened.

What makes it usable rather than merely automated is the discipline underneath. A summary may only contain what its source actually says — nothing is completed from the model’s general knowledge. Each edition covers one whole week, Sunday through Saturday, and consecutive editions tile exactly, so nothing is counted twice and nothing falls between two issues. Sources that stayed silent are listed as silent, because what was not covered is part of the picture. And the source list is published in full, each entry with a note on what it covers, how it reads, and where it needs reading against the grain — a vendor’s own claims marked as such, a broadcaster’s funding named, a channel whose framing leans one way said out loud. You can judge the digest because you can see what it read.

Beyond the weekly edition, everything accumulates: a searchable archive of every summary ever published on the topic, a feed, and optionally the same content as an email, a blog post or a narrated podcast — the pipeline writes to all of them from one run.

How I built it

A deterministic pipeline in TypeScript, not an autonomous agent: the same input takes the same path every time, and the language model is used only to transform text it was handed — summarize, condense, translate. Every branch, every ordering, every retry is ordinary code, which is what makes the output reproducible and the failure modes boring.

Each medium is one adapter behind a common interface, so a source is a line of configuration rather than a code change: YouTube via the Data API plus transcripts, podcasts via RSS plus speech-to-text, articles via feed or scrape, X via a search API. Adding a whole new topic is a single YAML file — sources, themes, prompts, schedule — which is why the marginal cost of the second and third topic is close to zero. Summaries land in a vector store as they are written, which is what makes the archive searchable later. Every model call is traced, so cost and quality per run are measurable rather than assumed.

Three topics run on it today: the AI-development digest it was generalized from — ~50 channels, weekly since 2025, public at aidevtube.com — a research-and-dealmaking digest built for the private-equity world, and a war-coverage digest reading Ukrainian, Russian, German, French and English sources side by side. The latter two are client- and access-controlled; the first is the one you can simply go and read.

Highlights

• Any topic: a new digest is a configuration file, not a project
• Multi-medium in one edition — video, audio, text and social read together
• Whole weeks, no overlap and no gaps, whatever time the run actually fires
• Full source transparency, including what each source is worth and what it isn’t
• Searchable archive of every summary, plus feed, email, blog and podcast output from one run

Stack

TypeScript · Node · PostgreSQL · Next.js · YouTube Data API · speech-to-text · Firecrawl · OpenRouter (LLM synthesis) · pgvector (searchable archive) · Langfuse (tracing) · Docker · Caddy