Loading…
Professional Services · United States
Grounded AI research assistant over 12M documents
A retrieval-grounded copilot cutting research prep time by 6 hours per analyst per week.
Client
Atlas Advisory (placeholder)
Services
AI & ML, Web Development
Duration
5 months
Year
2025
The challenge
Analysts spent hours per report searching decades of memos, decks, and transcripts stored across SharePoint, S3, and Notion. LLMs hallucinated citations. Compliance would not accept ungrounded answers.
Our solution
A retrieval-first system with strict citation policy: every claim links to source snippets, unattributable claims are refused, and an evaluation harness benchmarks every prompt change against a golden dataset.
How we approached it
01
Ingestion pipeline
Parsed 12M documents into semantically chunked, deduplicated vectors with source metadata preserved.
02
Citation-strict prompting
Structured outputs enforce that every claim carries a source ID — unattributable claims are rejected client-side.
03
Evaluation harness
A versioned golden dataset of 800 analyst questions gates every model or prompt change.
Technologies
Next.jsPythonpgvectorOpenAIAnthropic
Results
Outcomes measured against the goals we agreed to.
6
Analyst hours saved / wk
97%
Citation accuracy
92%
Adoption in 90 days
“The evaluation harness alone made this project worth it. We can now change prompts without holding our breath.”
Sofia Marín
Head of Research Technology · Atlas Advisory
More work
All case studiesFinTech2025
Consolidating $2.4B in cash operations onto one platform
Meridian replaced 14 spreadsheets and 3 legacy systems with a single treasury operating system for 40+ subsidiaries.
Healthcare2024
Building the operating system for a 60-clinic primary-care network
Helm replaced three point solutions with a unified operating system across 60 clinics.
Ready when you are
Have an idea? Let's build it.
Send us the shape of your problem. We'll reply within one business day with next steps.