How I built an open source News Discovery and Summarization feed, with Devin.ai, for $200
This site is the engineering log of discover-feed, a personal “Discover-style” news reader built almost entirely through sessions with Devin, a coding agent. Between October 7 and October 25, 2026, 44 pull requests — each opened by a Devin session from a written spec — took the repo from an empty scaffold to a running system: RSS/Atom and API ingestion, article clustering, LLM-written briefs and reports, web-search enrichment, context links, outlet-perspective synthesis, alerts, a public API with SSE, and a telemetry pipeline built to debug its own failures.
The app
The stack is boring on purpose: Python 3.12, FastAPI, SQLAlchemy over SQLite, APScheduler, and a server-rendered Jinja2 + HTMX + Tailwind UI. Summaries come from a local model — Qwen3.5 9B on llama.cpp, at roughly 42 tokens per second — behind an OpenAI-compatible endpoint, so any compatible endpoint can be swapped in. A story is a cluster of related articles; each story gets a brief (a lead plus cited bullet points) and, when it earns one, a longer report. Everything runs in Docker on the author’s own hardware; the deployed UI is reachable only on the author’s LAN.
The full feature list and the operating docs live on the repo’s documentation site — itself a Devin PR (#31).
How the work actually happened
Each session got a spec in the same shape: a context paragraph describing the repo, numbered tasks, acceptance criteria, non-goals, and a closing line — “open a PR and stop.” The agent ran the app locally, wrote tests, opened a PR, and the author reviewed, commented, and merged. Repeat 44 times.
That cadence is what this log is about. The chapters follow the pull requests:
- Chapter 1 — scaffolding, feed ingestion, deduplication, the first UI, text extraction and clustering.
- Chapter 2 — cited summaries, briefs and reports, the publish gate, the public API and SSE, story Q&A.
- Chapter 3 — exactly-once alerts, source management with budgets, SSRF-safe card images.
- Chapter 4 — web-search enrichment: adding other publishers’ coverage, hardening it, and tuning recall against content farms.
- Chapter 5 — background-and-context links, deep coverage for big stories, and per-source “how outlets differ” perspectives.
- Chapter 6 — observability: telemetry, the enrichment stall, the SQLite lock hunt, and making the pipeline visible.
Ground rules for this log
Every factual claim in these posts comes from one of three places: a merged pull request (always cited by number and linked), a file in the repo, or the author’s notes — reproduced verbatim where they’re used. No invented metrics, no invented quotes, no session transcripts, no dates that aren’t in the PR data. Where a spec’s description of the world turned out to be stale, the post says what actually happened instead.
The tone is deliberately a log, not a launch post. Several chapters exist because something failed — enrichment silently stopped for days, SQLite locks hit ~19 errors an hour, the instrumentation itself had a bug that made its most useful field unreadable. Those stories get told with the same specificity as the features, because they’re the parts that generalize.
How to read it
The timeline is generated from committed PR metadata and covers
all 44 PRs — including #28,
the one that never merged. Posts are grouped into six chapters in merge order;
each ends with a table of the PRs it covers (title, merge date, diff size), so
the claims are one click from their evidence. Tags cut across chapters
(enrichment, telemetry, debugging, docs, …) if you’d rather follow a
theme than a chronology.
This site is also itself a Devin session’s output — an Astro build generated
the same way everything else in the repo was, from a spec in the same format
it describes, deployed to Cloudflare Pages from the blog/ directory.