IN PRODUCTION · CASE ALIVE
Magellan Circle — a tender-intelligence agent in production
A tender-intelligence agent we built and run in production for Magellan Circle, an EU public-affairs and funding advisory. It ingests tenders from multiple public sources, enriches and retrieves against them, extracts structured criteria, and drafts fully-cited first-pass bid responses grounded in the client's own material — compressing weeks of manual research into hours.
- CLIENT
- Magellan Circle (EU public-affairs & funding advisory)
- DOMAIN
- Procurement / tender intelligence
- DURATION
- 12-month engagement
- SCOPE
- Multi-source EU tender ingestion · AI enrichment + semantic retrieval · structured-criteria extraction · retrieval-grounded, fully-cited first-pass bid drafting
- OUTCOME
- Weeks of manual tender research compressed to hours
- STATUS
- Live in production
- STACK
- Edge-native (Cloudflare) services · durable agent orchestration · hybrid + vector retrieval · eval + health monitoring · every claim cited (never-infer)
The problem
Magellan Circle advises clients on EU public funding and public affairs, and the raw material of that work is tenders — published continuously across multiple public sources, in inconsistent formats, with the qualifying criteria spread through long documents and annexes. Finding the opportunities worth pursuing, extracting what each one actually requires, and drafting a credible first response consumed weeks of skilled analyst time per cycle.
The bottleneck was structural, not a matter of effort: exhaustive reading does not scale with publication volume, and the judgement-heavy steps — what qualifies, what matches, what to say first — could not be reduced to fixed rules. It is the classic shape of an agentic workflow: the outcome is well defined, but the path varies with every tender.
The approach
We built an agent that runs the pipeline end to end. It ingests tenders daily from multiple public sources, enriches them, and retrieves against the client's own material using hybrid and vector search. For each tender it extracts the qualifying criteria into a structured form, assesses fit, and drafts a fully-cited first-pass bid response grounded in that retrieval — the agentic-RAG pattern, applied to procurement.
One rule defines the build: every claim in a draft is cited back to a source, and the agent never infers what it cannot ground. A run-time-chosen path is only trustworthy when its outputs can be checked, so checkability was designed in rather than reviewed in.
What makes it production-grade
The agent ships with the operational layer the demo never shows. An evaluation harness scores outputs against the bar agreed with the client, and health monitoring watches the live system — every run traced, so behaviour is diagnosed from evidence rather than guessed at. Orchestration is durable: ingestion and drafting survive source outages and retries without losing or duplicating work. The stack is edge-native on Cloudflare, which keeps the running footprint lean.
The outcome
Work that consumed weeks of manual research per cycle now takes hours. Analysts start from a structured assessment and a cited draft rather than a blank page — and because every claim is referenced, checking a draft is a review task, not a re-research task. Coverage changed shape too: the agent reads what its sources publish daily, so shortlisting starts from the full picture rather than from whatever a stretched team had time to sample. It has been live in production throughout a 12-month engagement, in daily use.