Case study · Axia Capital

Research that drafts itself.

A research agent for Axia's 28 analysts that ingests filings, transcripts, and sell-side notes, then drafts an internal memo with citations. 11 hours saved per analyst per week.

11 hrsSAVED PER ANALYST PER WEEK
94%CITATION ACCURACY
28ANALYSTS LIVE
3.4xRESEARCH OUTPUT
Engagement details
CLIENT
Axia Capital
INDUSTRY
Fintech · long/short hedge fund
DURATION
12 weeks · ongoing operate
TEAM
3 Mindlytic seniors + 4 Axia analysts
STARTED
Q4 2024
STATUS
Live · all 28 analysts onboarded
THE PROBLEM

Analysts spent 60% of their week reading.

Axia's analysts each covered 8–12 names across two sectors. The bottleneck was time — they spent the majority of every week reading 10-Ks, sell-side notes, and earnings transcripts, leaving only Friday for actual modeling and PM conversations.

Internal attempts at AI summarization had failed — outputs hallucinated numbers, missed risk-factor changes, and analysts couldn't trust them for client-facing memos.

WHAT WE BUILT

A research agent grounded in their own corpus.

We built a multi-step agent that ingests every relevant document for a covered company, runs targeted retrieval against Axia's proprietary corpus and SEC EDGAR, and produces a memo in the firm's house template — with every claim citing a source.

Critically, the agent refuses to make a claim if it can't ground it. Analysts review the draft, accept or reject claims, and the rejection signal feeds back into the eval set.

  • Ingestion: SEC EDGAR + S&P + Refinitiv + Axia's own note archive
  • Retrieval: hybrid BM25 + dense + cross-encoder rerank, per-analyst filters
  • Drafting: Claude 3.5 Sonnet with structured output and per-claim citations
  • Eval: golden memo set scored on factuality, completeness, and house-style
  • Feedback loop: analyst accepts/rejects feed weekly into the eval harness
WHAT WE FOUND

The trust problem was a citation problem.

Analysts didn't trust v1 because it gave them paragraphs with no way to verify. v2 changed every claim to a hover-able citation that opened the source PDF at the exact page. Trust scores doubled in a week.

We also found that 'completeness' mattered more than 'accuracy' — analysts could spot a wrong number, but they couldn't spot a missing one. We added a coverage rubric to the eval harness that scores whether every required risk factor and segment got addressed.

THE OUTCOME

11 hours back per analyst per week.

The agent now drafts the first version of every analyst memo at Axia. Analysts measure their time saved at 11 hours per week on average — used for more names covered (+3.4x research output) and more time with PMs.

Axia has since extended the engagement to cover their fund-level macro research and is piloting a similar system for their European desk.

They shipped in eight weeks what our internal team had been trying to build for fourteen months. And left us with a codebase we could actually maintain.
David Mbeki
CTO, Axia Capital

What we used.

LLM
Claude 3.5 Sonnet
Retrieval
Pinecone + BM25
Rerank
Cohere
Backend
FastAPI
Frontend
Next.js
DB
Postgres + S3
Infra
AWS Bedrock
Eval
Braintrust

Other case studies.

Have an AI project in mind?

Book a 30-minute discovery call. We'll tell you if we're the right fit — and if not, who is.