A working toolkit for private-equity firms who want to turn AI from deck-slide into measurable portfolio value. Twelve operations, all shipped on real public data — pandas math, row-level evidence on every dollar, MIT-licensed, air-gappable.
Most "AI for PE" today is slides, vendor demos, and a 100-day plan template no one operationalizes. Value gets left on the table at three layers — deal, portco, fund. Each layer has a partner-grade artifact missing on the desk. This stack ships those artifacts.
| Layer | What's broken today | What ships here |
|---|---|---|
| Deal team | DD memos hand-assembled; market scans take days; no consistent prospect scoring; CIM red-flag schedules built by associates over two-night reads. | /cim /explainer /exit-proof-pack /dd-checklist · plus a 17-command deal workbench in Claude Code |
| Portco operating | Operating partners can't compare decisions across companies; AI roadmaps drift inside a PDF; 100-day plans surface at the QBR, six weeks late. | /diagnose-decisions /plan-drift /normalize /procurement · top-3 dollar-quantified opportunities from the portco's own data |
| Fund / LP | No view of which portcos share the same leakage pattern; LP letters can't quantify operational alpha; DDQ packets contradict each other across vintages. | /benchmark-corpus /ddq /eval /ai-act-audit · fund-wide rank, archetype index, consistency-checked LP packets |
Every tool ships a static, auditable artifact a partner can defend. Every figure traces to a row in the underlying source data; every assumption sits in plain Python. Click Watch the demo to see the partner-grade walkthrough; Open the example to inspect the rendered report a portco would receive today.
Ingest a portco's CSVs; surface the top-3 repeatable, high-volume decisions being made badly — each with a counterfactual, a time-stability score, and row-level evidence. Pure pandas; the LLM never touches arithmetic.
Take any DX OpportunityMap and render a board-defendable memo — the why, the counterfactual, the risk of inaction, the rollout plan. Letterpress design; every figure traceable to a source field; whitelisted prose patterns.
Roll up N portcos into a fund-level rank table, archetype index, peer groups by cosine similarity, and quarter-over-quarter persistence — the LP-letter exhibit no scorecard vendor ships, because no scorecard vendor sees the row data.
The QA layer every PE-grade AI artifact needs. Citation accuracy, hallucination rate, coverage, consistency — four numbers, one rubric, deterministic scoring. Pure pandas plus regex; no LLM at runtime in the scoring layer.
Drop a CIM, an S-1, or any 10-K. Get back a section-cited list of red flags an associate would otherwise spend two nights surfacing by hand — eight heuristic flag families, every flag carrying a citation a reviewer can verify in fifteen seconds.
Buyer diligence will ask: prove the AI EBITDA. This is the trail you give them, before they ask, so the number doesn't get haircut at LOI. Provenance ledger, sensitivity table, defensibility checklist — schedule to the SPA, not the marketing deck.
The Q1 2026 ILPA AI diligence questions, answered against the actual artifacts in the fund's working directory — not against marketing copy. Every answer cites the file it came from. Cross-answer consistency checks run before the LP sees the packet.
Three portcos. Three different chart-of-accounts. One unified view, every cell traceable to its source field. The pre-step every cross-portco analysis needs and almost none of them do — defensible roll-up, every cell tracing home.
Day-Sixty of a hundred-day plan. The board is in seven weeks. This is the page the operating partner walks in with — diffed against real public 10-Q actuals, fetched from SEC EDGAR, no manual reconciliation. The page the consultant didn't ship.
Apollo's flagship value-creation lever, productized for the rest of mid-market PE. Public federal-contract data, no auth, no vendor. Same data Apollo's procurement team uses; same math; one-hundredth of the headcount.
Regulation (EU) 2024/1689. Article 6 high-risk classification. Deadline: 2 August 2026. Documentation skeleton, every obligation tracing to a public article — sized for the GC's red-pen pass, not for the law-firm memo.
The audit no fund runs on its own AI deployments and every fund should. Per-agent: model, run cost, run count, last-run date, health verdict — zombie, runaway, misaligned, or healthy. Three deterministic checks; no LLM at runtime.
Every operation lands on one role in the firm. Below is the partner-grade triage — the one tool to open in your first thirty minutes with this stack.
/diagnose-decisions/cim /explainer/benchmark-corpus/ddq /eval/diagnose-decisionsThe artifact a partner can defend is built by the math underneath, not the prose on top. Every tool in the workshop honors the same four contracts.