A deal-sourcing and screening dashboard built around a specific, opinionated thesis: PE doesn't play well at the AI labs/compute layer — it plays in the fragmented, already-profitable SMB service businesses (bookkeeping, IT support, medical billing, home services) where AI is a margin lever on an existing cash-flow base, not the product itself.
All data is fictional. Every company, financial figure, and deal note is solely illustrative — built to demonstrate a sourcing/screening framework end-to-end, not to represent real targets.
Live: https://bakulbadwal.github.io/dealdocket/
Open it: clone the repo and run a local static server (see below) — the app fetches data.json, which browsers block over file://.
Screening
- The five-box screen — five weighted criteria (market fragmentation, unit economics, AI-adoption leverage, moat & stickiness, exit path), each with a live slider. Drag any weight and the entire 30-deal pipeline re-ranks in real time — the point is to make the thesis's sensitivity to its own assumptions visible, not just show a static scorecard.
- Weight presets — one-click lens switches (Balanced, AI-Leverage Max, Margin First, Moat & Exit) that jump all five sliders at once; any manual drag hands control back to you.
- Rank stability — the sharp question about any weighted score is which results survive a change of mind. Every deal is ranked across 2,000 weightings sampled uniformly from the simplex, and the drawer reports its best/median/worst rank, its p10–p90 band, and the share of weightings where it lands top-five. A narrow band means the deal is genuinely strong; a wide one means its rank is an artefact of how you happen to be weighting today. Seeded, so the figures are identical on every load.
- A 30-deal illustrative pipeline across 10 service verticals, spanning every stage from sourced to closed to passed — including deals that fail the screen, not just wins.
Modelling the thesis
- The roll-up builder — screening ranks targets but never shows why the strategy makes money. Pick a platform and its add-ons and the tool models the combination: pro-forma revenue, EBITDA, margin and headcount; the blended entry multiple; and a value bridge that separates the two real sources of return — multiple arbitrage (buy sub-scale, exit at the platform multiple, no operating change) from AI margin expansion (each vertical's automatable cost base × how much you capture × that asset's own AI-leverage score). The bridge is constructed to close exactly:
entry + arbitrage + AI + organic === exit value. - Overlaid profiles — the radars of every selected asset drawn on one chart, answering the question you actually have while assembling a roll-up: are these the same shape, or complementary? A roll-up of identical profiles concentrates the same risk.
- Live assumptions — AI cost capture, organic growth and exit multiple are all sliders, and the tool states its own limitations next to the output (all-equity, pre-fee, no integration cost, no failed add-on).
Using it
- Shareable state — weights, filters, sort and the entire roll-up selection are encoded in the URL, so a link reproduces a specific reading rather than the default view. Mirrored to
localStorage, so returning to the page restores your last calibration. - Deal detail drawer — per-criterion rationale, financials, ask and entry multiple, headcount, recurring-revenue share, and the rank-stability band.
- Filters and live stats — search, vertical/stage/channel filters, active-pipeline-only toggle, and a stats bar.
- Keyboard accessible — every deal is a real focusable control, the drawer is a proper dialog with focus trapping and return, and there's a visible focus ring throughout.
The visual language is deliberately terminal-inspired — monospace numerals, hairline rules, a single green accent that carries meaning (high scores and healthy stages) rather than decoration.
- CASE_STUDY.md — the product write-up: the problem being modelled, who it's for, the decisions and tradeoffs, and how I'd measure whether it works.
- ROADMAP.md — what this doesn't do yet and why, including the honest weak spot (the automatable-cost estimates driving the AI uplift are mine, not researched — and are tagged as such in the data), plus what's deliberately not being built.
Financials are stored as numbers, not display strings, so every figure on the page is computed rather than written down. Entry multiples are a deterministic function of deal size and the moat/exit-path scores — reproducible, not random. The margin is derived from revenue and EBITDA at render time so it cannot drift out of sync with them.
The one input that is not derived is automatableCostShare per vertical, which drives the AI uplift. Those are estimates, marked "basis": "estimate" in data.json with a null source. See the roadmap.
Same data/view split as the AI Stack: data.json holds every deal, the framework definition, and the thesis copy; app.js renders it and runs the scoring engine; styles.css handles presentation. A sourcing pipeline is exactly the kind of content that should be able to update independently of the render logic.
git clone https://github.com/bakulbadwal/dealdocket.git
cd dealdocket
python3 -m http.server 8000
Then open http://localhost:8000/.
Vanilla HTML/CSS/JS. No framework, no build step, no dependencies.
MIT — see LICENSE.