FastOrchestra
An application pipeline over ranked job ads
A harvest runs every day, reads multiple job boards and ranks what it finds. FastOrchestra takes those ranked rows and keeps everything that happens next in a store of its own: the shortlist, the stage each application sits at, the dated evidence behind it, and the letters. Each letter is drafted from the profile, the ad, and the sent letters nearest to it, and every model call is traced.
A 38-second tour. The longer cut takes four minutes and walks through the whole architecture.
What It Does
Ranked Ads with Fit and Gaps
All ads are ranked during a scheduled process. The interface presents each ranked ad with its fit and gap indicators, highlights semantically closest postings, and marks potential duplicates.
A Shortlist with Stages
A shortlisted ad becomes an application on a board, one column per stage. Every move is dated, and the evidence that justified it stays attached to it.
Letters Drafted from Context
A draft is assembled from the profile, the ad itself, and the letters of comparable applications already sent — retrieved by meaning, not by keyword. The examples are your own sent mail, so they grow with every application. Versions are kept, and the current one is the one that went out.
Every Model Call Traced
Prompt, reply, tokens, latency and cost land in Langfuse under one trace id. Sync staleness, embedding coverage and draft spend are gauges Prometheus scrapes.
How It Is Built
- Daily harvestmultiple boards, in n8n
- Rankan LLM scores each ad 0 to 10
- One storePostgres 16 with pgvector
- Draftone call leaves the host
- Traceprompt, tokens, latency, cost
React 19 and TypeScript at the front, FastAPI and SQLAlchemy 2 behind them, one image running two processes: an API, and a worker that drains a queue held in the same store. Ads and the profile are embedded on the host, on CPU, so only the drafting prompt leaves it. The API also speaks the Model Context Protocol, so an assistant can list the ranked ads, shortlist one, and submit the letter it wrote.
Want a Pipeline Like This?
Ranked inputs, a store you own, model calls you can audit after the fact, and an assistant that can drive the whole thing. The same shape fits lead triage, document review, or routing what arrives in a shared inbox. Tell us what yours has to do.
Get in Touch