Agentic & Multi-Agentic Pipelines
Systems built from multiple specialized AI agents coordinating toward one outcome.
Cold War: Decided
An evidence-disciplined, production-deployed multi-agent decision simulator — document ingestion, specialist agents, independent review, deterministic guardrails, human authorization, and a persistent audit trail.
Four evidence-siloed specialists (Submarine, ELINT, Air, HUMINT) analyze the same crisis from isolated dossiers, get targeted Chief feedback, revise once each, and pass through Counterintelligence red-team review before a Decision Auditor challenges the proposal and a deterministic software policy matrix — not free-form prose — decides which actions are actually eligible. A human still authorizes every outcome. The system ingests PDF scenarios with page-level citations, and every one of the 16 pipeline stages is preserved in a persistent Postgres mission ledger for later inspection. The Cold War setting is the user-facing theme; the reusable part is the architecture underneath.
Opportunities to learn
- Having built a rule-based evaluator to score evidence discipline and independence, a future project could pit it against an LLM-as-judge grader on the same transcripts and see where they disagree
- Now that every mission stage lands in an immutable ledger, a future project could add a side-by-side replay view to diff two runs of the same scenario
Blackstone Carry Review
Six Claude agents turn public Blackstone carry disclosures into a graded review memo, while a separate calculator keeps the waterfall math in ordinary JavaScript.
Built to test what happens when one AI agent's output becomes the next agent's only source. A retriever pulls SEC and Blackstone material; a mechanics agent explains hurdles, catch-up, and the GP/LP split; a summarizer compresses it; a skeptic challenges what survived; and a final judge grades the memo against the original source. A separate branch handles the numbers differently: normal JavaScript calculates a hypothetical fund waterfall, then Claude explains the already-computed result in plain English instead of doing the math itself. Repeated testing caught both one-off agent drift and a real timing bug in the calculator. The underlying fund table is a dated teaching dataset, so the tool is intentionally presented as a learning experiment — not an estimate of Blackstone's actual carry.
Opportunities to learn
- Having worked with a dated teaching table here, a future version could rebuild the inputs from versioned, independently verified source records with explicit provenance
- Now that the six-stage chain is observable, a future project could compare repeated runs and measure which handoffs most often lose or distort facts