Equity Research Terminal
Multi-modal RAG over 10 years of filings, transcripts, and broker research with citation-grounded answers.
- Client
- Atlas Capital
- Year
- 2025
- Duration
- 12 weeks
- Team
- 3 engineers, 1 strategist
Atlas Capital's equity analysts were spending roughly 60% of their day searching rather than analyzing. We built a research terminal over a decade of filings, earnings transcripts, and licensed broker research — with span-level citations on every claim and information-barrier rules enforced inside the retrieval query rather than after generation.
The brief
Give analysts a single research surface that compliance would approve, where every answer is traceable to the document and page it came from.
- Finance
- RAG
- Citations
The decisions that mattered.
Every engagement turns on a handful of calls. These are the ones that decided whether this system reached production.
The benchmark came before the system
We built a 1,200-question labeled benchmark with Atlas analysts before writing retrieval code. Every subsequent decision was justified by whether it moved recall@10 on that set.
Hybrid retrieval, because vectors miss tickers
Pure dense retrieval reliably failed on exact identifiers, CUSIPs, and line-item labels. Combining BM25 with embeddings and a cross-encoder reranker lifted recall@10 from 71% to 94%.
Information barriers enforced at query time
Wall assignments filter the candidate set before generation, so restricted material cannot influence an answer even indirectly. This was the condition compliance set for approval.
Groundedness checked on every answer
A verification pass confirms each claim is supported by a retrieved span. Unsupported statements are suppressed rather than shown with a weak citation.
What shipped, and what changed.
Outcomes
- Analyst throughput doubled on coverage tasks
- Answer grounding rate of 97% on the held-out benchmark
- New-analyst onboarding time cut in half
- Zero information-barrier exceptions across the first year
“Compliance approved it in one review cycle. In this firm, that is the strongest endorsement a system can get.”
What we built it with.
- Claude
- pgvector
- Elasticsearch
- Python
- LangChain
- Next.js
Other projects from the shelf.
Clinical Documentation Copilot
Ambient AI scribe that drafts structured notes, ICD-10 codes, and prior-auth letters in real time.
Read morePredictive Maintenance Platform
Time-series + vision models predicting equipment failure 72 hours in advance across 14 plants.
Read moreAI Concierge for Retail
End-to-end conversational agent resolving 62% of post-purchase requests without human escalation.
Read moreTell us about the problem.
A 30-minute call with the senior team. We will tell you what is realistic, what it would cost, and whether we are the right people for it.