Grounding an AI research analyst on a decade of data
Analyst throughput doubled on coverage tasks and onboarding time for new hires dropped by half.
- Industry
- Financial services
- Size
- $18B AUM, 140 analysts
- Engagement
- 12 weeks
- Team
- 3 engineers, 1 strategist
- Analyst throughput
- 2xAnalyst throughput
- Onboarding time
- −50%Onboarding time
- Answer grounding
- 97%Answer grounding
Analysts spent 60% of their day searching filings, transcripts, and broker notes for context.
Atlas analysts estimated they spent 60% of their day searching rather than analyzing — across a decade of filings, earnings transcripts, and licensed broker research spread over four separate systems. Two prior vendor pilots had failed compliance review: neither could demonstrate where an answer came from, and neither enforced the firm's information barriers at retrieval time.
The engagement
- Industry
- Financial services
- Size
- $18B AUM, 140 analysts
- Engagement
- 12 weeks
- Team
- 3 engineers, 1 strategist
A citation-grounded RAG system with role-aware controls, evaluated against a 1,200-question benchmark.
Four phases, each with a written definition of done.
Build the benchmark first
Before any retrieval code, we worked with Atlas analysts to build a 1,200-question labeled benchmark. Every subsequent decision had to move recall@10 on that set or it did not ship.
Fix ingestion before retrieval
Most early failures traced back to parsing, not search. Structure-aware chunking of tables and financial statements moved the number more than any embedding change.
Hybrid retrieval with reranking
Dense embeddings plus BM25 with a cross-encoder reranker lifted recall@10 from 71% to 94%. Pure vector search consistently missed exact tickers and line-item labels.
Enforce barriers at query time
Wall assignments filter the candidate set before generation, so restricted material cannot influence an answer. This was compliance's condition for approval.
What changed.
Analyst throughput doubled on coverage tasks and onboarding time for new hires dropped by half — new analysts could answer questions on unfamiliar names without first learning where everything lived. Compliance approved the system in a single review cycle, and the firm has recorded no information-barrier exceptions in the year since launch.
- Analyst throughput
- 2xAnalyst throughput
- Onboarding time
- −50%Onboarding time
- Answer grounding
- 97%Answer grounding
“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
The capabilities behind this engagement.
More of our work in finance.
Risk, fraud, AML, and advisor copilots grounded in regulatory frameworks and the documents that drive them.
Other engagements worth reading.
Let's talk about what this would look like for you.
Thirty minutes with the senior team. Bring the problem and we will give you an honest read on scope, timeline, and whether we are the right fit.