We turned 200 daily alerts into one 2-minute briefing (with LLMs)

We're building EveBrief — a tool that watches a set of sources you care about and sends you one readable briefing instead of a stream of alerts. This is how the pipeline works, what it costs, and the parts that are genuinely hard.

Pre-launch, building in public. If you want the briefing instead of the build log, the waitlist is the place.

The input problem

A realistic monitoring setup for one small agency looks like:

• 12 RSS feeds (competitor blogs, two trade publications, a subreddit or two)
• 6 Google News queries ("[competitor] pricing", "[competitor] funding", …)
• 8 URLs with no feed — pricing pages, changelogs, docs — that we poll for content changes
• a few keyword alerts

On a normal day that's 150–250 raw items. Maybe 5–10 of them matter. The other ~95% is reposts, listicles, press-release syndication, and the same story rewritten by eight outlets.

The job is to get from 200 items to a briefing a busy person will actually read in two minutes — grouped, deduplicated, and with a line on why each thing matters to them specifically.

The pipeline

Four stages. Nothing exotic.

1. Collect

Feeds and Google News come in as RSS — cheap, reliable, well-structured. For the no-feed URLs we fetch the page, extract the main content, and diff it against the last snapshot; a meaningful change becomes an "item" with a before/after.

Everything normalises to the same shape: {title, url, source, published, text}. We pull the first few hundred words of body text where we can get it — the LLM step is much better with a paragraph than with a headline.

2. Deduplicate

This is where most of the 200 → 30 reduction happens, and it's the least glamorous part.

• Exact URL and near-URL matches (tracking-param stripping) — easy.
• Same-story-different-outlet: we cluster by title similarity plus a cheap embedding on the first paragraph, then keep the most authoritative / earliest member of each cluster and attach the rest as "also covered by".
• Evergreen filtering: a lot of "news" search results are actually docs pages, wiki entries, and vendor comparison pages that rank forever. Undated or old-dated items with generic titles get dropped.

After this stage you're usually at 20–40 candidate items.

3. Filter and rank against stated interest

The user has written, in plain English, what they care about: "pricing changes, new features, funding, and key hires at these four competitors; anything about AI Overviews."

We turn that into a relevance pass: each candidate gets scored for how well it matches the stated interests, with a hard requirement that a specific subject term actually appears (so "governance" or "revenue" alone doesn't sneak a generic item through). Low scorers are dropped and logged as "filtered as noise" so the user can see what was thrown away and trust the filter.

Down to 5–12 items.

4. Synthesise the briefing

One LLM pass over the surviving items plus the reader profile. Output is structured: theme groups, each item with a two-line TL;DR and a "why it matters to you" written against the stated interests, plus a short "what the filter dropped" line.

The prompt does a lot of unglamorous work: forbidding hype language, forcing observational phrasing over instructions to the reader, keeping the "why it matters" concrete rather than "this is significant for the industry."

What it costs

Per briefing, for one project:

• Collection: RSS is free; change-detection fetches are a few hundred HTTP requests. Negligible.
• Dedup embeddings: ~30–40 short embeddings. Fractions of a cent.
• Synthesis: one call, ~4–8k tokens in, ~1–2k out.

On a small hosted model that's well under a cent per briefing. On a frontier model, a few cents. Either way the cost driver is not the LLM — it's building and maintaining the collectors.

The parts that are actually hard

Source coverage. The LLM step is easy. Getting comprehensive, low-noise input is the whole game, and it's endless: feeds move, sites add bot protection, a "news" search silently starts returning SEO pages.

Dedup precision vs recall. Cluster too aggressively and you merge two genuinely different stories. Too loosely and the same acquisition shows up three times. We bias toward keeping items separate and merging visibly.

"Why it matters" without hallucinating. The valuable sentence is the one that connects a news item to the reader's situation. It's also the easiest place for a model to invent a plausible-sounding connection. We keep that sentence tied to the stated interests and the item's own text, and it still needs a human eye before anything goes out at launch.

Knowing what you dropped. A filter you can't inspect is a filter you won't trust. Showing the "filtered as noise" list — and getting it right — matters as much as the briefing itself.

Where this is

The pipeline runs and produces real briefings — there's a full example on the site, generated from real sourced news for a stand-in agency. We're validating whether people want this enough to pay for it before building the full self-serve product.

If "one briefing, not 200 alerts" is a thing you'd use: join the waitlist.