Scout
An AI tool that replaces a team of financial analysts — built for a single-person PE firm.
Most funds spend 70 % of their time on screening. I wanted to spend 20 %. So I built Scout — a one-click pipeline that pulls Czech corporate filings, extracts financials, and scores companies against Point FM’s thesis.
One click on a company runs a four-step pipeline.
1. Pulls filings from Justice.cz Sbírka listin — preferring complete annual reports, ignoring notarial filings and audit reports.
2. Walks back year-by-year until it has three readable years of financials.
3. Sends PDFs and XMLs (including scans) to Claude Sonnet 4.6 — extracts revenue, EBITDA, net income, assets, equity, headcount.
4. Computes PFM thesis scores, runs DCF, applies sector multiples, returns a fair-value range.
Manual version: 15–20 minutes per company. Scout: 30–90 seconds. A 200-company sector screen used to take three people three weeks. Now I have it by next morning.
Pulled seven years of filings (2017–2024). Extracted full financial series. Computed PFM scores: Growth 57, Margin 93, Efficiency 95, Stability 60 — Strong Fit against the thesis. DCF with WACC 12.37 % and terminal growth 2.5 %, plus C2C marketplace sector multiples (25 % minority discount), gave a fair-value equity range of CZK 613M – 1B, point estimate CZK 830M.
The 16 % gap between DCF and multiples is itself a signal — peer multiples capture strategic premia (synergies, scale, multi-year visibility) that DCF can’t. Scout flags it and leaves the call to a human.
The most interesting thing it produced unprompted, sourced from its own Knowledge Base:
“AI/agent disruption — purchasing agents may commoditize the browsing/discovery layer of marketplaces in 3–5 years.”
That’s a key risk flag I never put in the prompt. Same for the headcount drop from 40 to 26 while revenue grew, and the missing 2022 data.
GMV, take rate, unit economics, churn, product roadmap, portfolio synergies. Only the owner sees those.
Scout is a first-pass screening tool, not a final valuation. Insider valuation will always differ. It’s a wide net, not a substitute for ownership-level judgment.
Next.js 16, Supabase (Postgres + pgvector), Claude Sonnet 4.6 for extraction and analysis, OpenAI embeddings for the Knowledge Base, Inngest for background jobs, Vercel hosting. Public sources: ARES v3 and Justice.cz Sbírka listin. Consciously minimal stack — no custom backend.
Currently testing connections to external Polish and Austrian private-company databases. Built with Claude Code as a pair programmer — a tool for AI-driven investing, built by AI-driven coding.
“The first screening of a sector with 200 companies used to take three people three weeks. Now I have it by next morning. It’s not about speed — it’s about how much of the market I can cover as one person.”
Scout was built for Point FM. A few PE firms have asked about access. I haven’t worked out a business model — if you’d find it useful, let’s talk.