Read the case study for the business story and results.
Sourcing and diligence shared the investment profile.
The build combined a lookalike sourcing model with a diligence co-pilot. Partner interviews, historical investments, and acquired-founder feedback informed the firm's investment profile.
The co-pilot worked with private-company data rooms and commercial research. It planned the task, retrieved supporting extracts, and composed source-linked memos. The deployment used governed AWS cloud infrastructure with an in-VPC retrieval index and controls separating confidential deal material.
Historical information kept the sourcing test honest.
The test used a fund-boundary cutoff and signals available at that time. Training used earlier investments; evaluation ranked ~5,000 eligible businesses and checked where eight later acquisitions appeared.
Six ranked in the top 200 and all eight in the top quarter. In acquired-versus-declined comparisons, the acquired business ranked higher 82% of the time. These tests measured alignment with the firm's choices, not future investment returns.
We evaluated retrieval and claim support separately.
On 250 analyst questions, retrieval recall among the top ten results was 92.2%. Claim-level groundedness was 99.2% across 1,500 claims: the share supported by their source evidence.
Two partners reviewed eight paired memos blind and found no critical errors. Source-pinning, abstention, and mandatory partner review remained in the production workflow. These are results from defined evaluation sets, not a guarantee that future outputs can't be wrong.
We would increase the claim sample in a future validation. A strong observed percentage still benefits from broader coverage of the questions and documents the team will encounter.
Management red flags needed a second gate.
A same-name mismatch tied an adverse item to the wrong executive. We added an entity-match confidence threshold, stronger cross-source identity checks, and mandatory human approval before a management red flag entered the memo.
Review stayed with the investment team.
A designated champion introduced the pipeline in weekly meetings. We trained the team to verify source claims rather than accept a well-written memo at face value. Analysts and partners retained the valuation model and investment decision.
The ~$22M figure is the value of the sourced deal the firm closed. Operating-cost reductions and deal value remain separate outcomes. The eight-week delivery covered sourcing and diligence; it wasn't a portfolio-company transformation.
The deal log measured evaluations per analyst per period. An evaluation ran from first awareness of a business to the internal final memo. Each analyst completed 4× as many evaluations; with the team moving from seven to four, that same measurement implies 4 × 4/7 = ~2.29× total throughput. These are two views of one productivity result.