CLEAR
A method for building committees of AI experts — proven out on investing first
Take whatever actually captures how a specific expert thinks — a public figure’s books and interviews, or just as easily a consultant’s confidential reports, or years of internal notes from an expert inside your own company, still there or long retired — and run it through a method I call Philosophy Server Technology: a way of turning that record into an independent AI expert with its own coherent point of view, not a chatbot doing an impression of someone. Do that for several people whose philosophies genuinely disagree with each other, and you don’t end up with more chatbots — you end up with a committee: real humans on your team, working alongside simulated experts who amplify what that team is actually capable of, instead of one model’s best guess dressed up as an answer.
Say the problem is cutting shipping costs and improving on-time delivery across a global supply chain. There’s no single right answer — there are maybe five people alive with genuinely different, defensible philosophies about how to solve it, and the honest move is to put your specific situation in front of all five and see where they agree and where they clash. That’s what the diagram above is describing in the abstract.
The same idea holds anywhere a single generic assistant hands you plausible-sounding output instead of a real, defensible answer: legal review, medical triage, engineering sign-off, hiring, product strategy. It’s a general pattern, not a finance trick — investing is just the domain where I’ve built it all the way out, top to bottom, so far.
The application is called CLEAR, and its committee is five investors, chosen because their philosophies genuinely conflict: Warren Buffett, Charlie Munger, Seth Klarman, and Peter Lynch each cover a different flavor of value or growth investing — and because none of those four really touch quantitative, market-neutral thinking, the fifth seat belongs to Jim Simons.
Five independent evaluators, not five hats on one model
Each of the five is its own evaluator — not a shared prompt with a name swapped in. Buffett’s evaluator carries real disqualifying logic: circle-of-competence checks, moat-strength scoring, net-debt-to-EBITDA thresholds, even bank-specific CET1 capital-ratio checks when the target is a financial. Nearly 1,800 lines of that logic exist because the difference between Buffett’s discipline and Klarman’s isn’t a matter of tone — it’s a matter of what disqualifies an idea before you ever get to scoring it.
This runs on a real stack, not a demo: FastAPI backend, a React/TypeScript frontend (with English, German, Spanish, and Italian support built in), kdb+ ingesting SEC XBRL filings directly, Docker and systemd in production, CI/CD behind it. There’s an internal RBAC design, a documented rollback/migration runbook, and a real test suite — the boring infrastructure nobody puts in the demo, which is exactly why most “AI stock picker” side projects never have it.
The part that’s actually novel
Two things in here go beyond “LLM plus finance app”:
Chat with the Gurus. Each philosophy persona runs as its own isolated Claude Code subprocess with its own conversation context, tied directly to that evaluation’s data — so when you push back on the Buffett read, you’re arguing with the Buffett evaluator’s actual reasoning, not a generic assistant improvising.
A financial-advisor MCP server. CLEAR’s own analysis engine is exposed as a set of Claude Code tools — analyze_stock, compare_stocks, and more — so the philosophy logic isn’t trapped inside a web app. It’s callable infrastructure.
A semantic consistency pipeline that scales. Every variable definition — a formula, a threshold, a range — has to stay consistent across the glossary, the Python scoring code, the TypeScript types, the per-philosophy rationale text, the tooltips, even video scripts. Checking that by hand is a multi-hour slog per pass. CLEAR uses Claude Code itself as the semantic comparator across all six-plus sources, with a batch mode that runs the full 45-variable check in 30–60 minutes instead of the 5+ hours a naive per-variable pass takes. That’s not a finance feature. That’s a pattern for keeping any complex system’s stated logic and its running code honest — automatically, on a schedule, instead of hoping someone notices the drift.
5 independent expert evaluators, each with its own disqualifying logic and scoring model
10–20x faster consistency verification across code, docs, and UI via batched semantic diffing
4 languages shipped in the UI, because production means production
Why this matters beyond investing
Investing is just where the committee happens to be finished. The pattern underneath it — independent AI experts, each holding a distinct, coherent philosophy, working alongside a human instead of replacing their judgment — doesn’t care what the domain is. Anywhere a decision is genuinely contested, and the honest answer is “it depends who you ask,” a committee beats a single assistant. CLEAR is the proof that this can be built for real, not just pitched.
Have a problem that actually needs a range of expert opinion, not one model’s best guess? Reply and tell me what you’re deciding. I’ll tell you straight whether the pattern fits.