A small foundation team we work with needs non-dilutive funding — grants, partnerships, corporate CSR — across many countries and programmes. Done by hand, that means tracking hundreds of funders, watching the news for new calls, researching who decides, judging which opportunities are worth pursuing, and writing formal applications in two languages. It doesn’t scale, and the best-fit opportunities slip past while someone is buried in a spreadsheet.
So we built the pipeline that does the legwork — and stops, deliberately, before anything is sent.
A chain of agents, each with one job
The system is a lead lifecycle — prospect → researching → drafting → review → submitted → decision — driven by a chain of autonomous agents, each with a defined contract, their work streaming to the operator’s screen in real time:
- Signal Monitor — a daily sweep of news and funding feeds, triaged by an LLM for genuine relevance.
- Funder Prospector — discovers new opportunities via research APIs and de-duplicates them against the existing catalogue.
- Contact Hunter — finds the actual decision-maker, with a fallback chain when the obvious source comes up empty.
- Fit Scorer — scores each opportunity 0–100 on weighted dimensions, with a written rationale — so the team spends its hours on the leads worth pursuing.
- Outreach Drafter — generates the formal deliverables (concept note, full application, one-pager, cover email) in German or English, tuned to real application structure, with versioning.
The most important part is where it stops
Everything above is automation. The design decision that makes it trustworthy is the human-in-the-loop gate: no document leaves the building until a person has reviewed it, approved it, and explicitly clicked send. Two steps, on purpose. The agents do the exhausting 90% — discovery, research, scoring, a solid first draft — and hand a human the 10% that actually needs judgement.
Under the hood it’s built for the real world: a resilient four-provider LLM cascade with per-service quota gating and live cost tracking, real-time streaming for long-running agents, and a multi-tenant design so one team can run the same engine across several partner projects, each ranked against a shared catalogue of opportunities.
Lead-gen is a shape, not a sector
We built this for fundraising, but the shape is general: monitor signals → score fit → find the contact → draft the outreach → gate on a human. It’s the same pattern behind B2B lead generation, partner discovery and business development anywhere the work is high-volume research plus careful, personalised writing.
The goal was never to automate the judgement. It was to clear everything out of the way until the judgement — and keep a human firmly on the trigger.