Capital Committee
Engineering economics & case practice. Put a capital decision in front of a simulated investment committee of seven Claude-powered departments, model it with NPV / IRR / tornado sensitivity, and drill consulting case interviews, all grounded in sourced GCC and industry figures.
Live demo · Case study · Designed & built by Ayman Khayat
The live demo runs in demo mode on GitHub Pages: sample committees replay, and the calculator and case practice work fully. Live AI runs need the serverless API on Vercel with an ANTHROPIC_API_KEY (see Deploy).
The hero and boardroom images are concept images made with Higgsfield from the project's real renders, and are labelled as such on the site. Prompts are logged in docs/higgsfield_prompts.md.
Three modes
1. Company Simulation (main page)
The page opens on a rounded hero panel, followed by a Live Committee section holding the real app: a 3D committee map (three.js) of metallic department hubs orbiting the Company Brain, with a live particle core, drifting dust and beads of light flowing along each department's link. Hovering a department flies the camera in and sets off a burst of its colour. Clicking zooms further and reveals that department's four specialist agents; once a verdict is in, the panel shows what each agent contributed. Without WebGL, or with reduced motion requested, it falls back to the flat SVG map. The render loop pauses while the scene is off-screen.
A decision runs through seven department agents, in parallel, each a separate Claude call with its own persona:
| Department | Argues from | Characteristic bias (kept on purpose) |
|---|
| Finance | NPV/IRR vs hurdle, payback, covenant headroom | Underweights option value and long horizons |
| Strategy & Corp Dev | Right to win, timing, build/borrow/buy | Optimistic about upside; treats execution as detail |
| Legal & Risk | Exposure, conditions precedent, RAPID "Agree" veto | Defensibility over expected value |
| Operations & Engineering | FEL stage, AACE estimate class, HSE, critical path | Wants one more stage of definition |
| Marketing & Sales | Bottom-up SOM, win rates, LTV:CAC | Optimistic demand, light on cost-to-serve |
| Sustainability & ESG | GHG scopes, SBTi / SGI / UAE Net Zero, IFRS S2 | Underweights near-term cash |
| HR & Talent | Capability gaps, Nitaqat / Emiratisation, ADKAR | "The organisation isn't ready" |
Each returns a schema-validated verdict (stance, confidence, 2–3 reasons, one risk flag, conditions, key metric). Only after all seven settle does the Company Brain rule. It acts as the RAPID "D", honours Legal and safety vetoes, corrects for each function's known bias, names the single hinge trade-off, and sets owned conditions and kill criteria.
- Boardroom ↔ Quick verdict views
- Culture weight sliders (finance-driven, engineering-driven, sustainability-first, growth-led). The weighted lean is computed deterministically and updates live; re-convening the Brain is one extra call.
- Show disagreement draws the conflicts on the map and filters the cards to departments that clash
- "Your call" panel. The Brain always rules, but flags close high-severity conflicts that turn on risk appetite so a human can overrule it
- Departments can be switched off per run; IC memo export (print to PDF); run history kept in the browser
2. Calculator
An assumptions sheet (blue = input, financial-model convention), a live WACC slider with GCC reference bands, and NPV, IRR, MIRR, simple and discounted payback, PI and LCOE. Four exhibits in consulting style, each with a "so what" action title, footnotes and a source line:
- Tornado sorted by swing, engineering low/high ranges (toggle to ±20%), coloured by input level, not outcome
- NPV profile against discount rate, with the IRR crossing marked
- Breakeven table: how far each input can move before NPV = 0
- Cash-flow schedule
Presets: 10 MWp factory solar (KSA), 5 MWe waste-heat-to-power (cement), SWRO energy-recovery retrofit (UAE), and a deliberately negative battery-arbitrage case. Send to the committee attaches the results as a committee pack, so Finance argues from your numbers.
3. Case Practice
16 original cases: 8 strategy (market sizing, profitability, breakeven, pricing, market entry, M&A) and 8 technical (energy go/no-go, LCOE/LCOW, debottlenecking, district cooling, plant efficiency). 12 are set in the Gulf.
- Entry-level ↔ advanced toggle, calibrated against public casebooks; each case is tagged with its difficulty drivers (trap, multi-stage, second-order effects…) and interview format (interviewer-led, candidate-led, engineering consulting)
- Timer, exhibits you request one at a time, a framework hint that nudges without giving the answer, then a reveal of the model issue tree, reasoning, answer and the partner-style synthesis
- Self-scoring on the four-part rubric, plus optional AI interviewer feedback
- Progress history (attempts, rubric averages, frameworks practised) persists in the browser
Architecture
flowchart LR
B[Browser] -- 7 parallel POST /api/department --> F1[Serverless fn × 7]
F1 -- Claude API --> C[(claude-opus-5 / claude-sonnet-5)]
B -- after all settle: POST /api/brain --> F2[Serverless fn]
F2 -- Claude API --> C
B -. static .-> S[index.html + ES modules]
- The API key never reaches the browser. Vercel serverless functions in
/api hold ANTHROPIC_API_KEY and proxy every call. An optional APP_PASSCODE protects your budget on a public URL, and a best-effort per-IP limiter caps abuse.
- Parallelism happens in the browser. Seven independent function invocations, so cards land as each department finishes and wall-clock time is roughly one call plus the Brain, not eight in sequence.
- Structured outputs (
output_config.format with JSON Schema) guarantee parseable verdicts; adaptive thinking; server-side refusal fallback on Opus 5; the persona system prompts are prompt-cached.
- No build step. Vanilla ES modules and hand-drawn SVG charts, plus one dependency (
@anthropic-ai/sdk).
- Without a key the site runs in demo mode and replays sample committees for two decisions.
Cost per run (list prices, Sep 2026)
| Tier | Departments | Brain | Typical full run |
|---|
| Standard | Opus 5 ($5 / $25 per M tokens) | Opus 5 | ≈ $0.40 |
| Economy | Sonnet 5 ($2 / $10) | Opus 5 | ≈ $0.20 |
Roughly 1.3K input and 1.5K output tokens per department (including adaptive thinking) and about 5K / 2.5K for the Brain. The header shows the actual spend from each response's usage.
Run locally
npm install
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env.local
npm run dev # http://localhost:5173
npm test # finance engine checks against the research sanity figures
Deploy (Vercel)
Import the repo in Vercel, add ANTHROPIC_API_KEY (and optionally APP_PASSCODE) under Environment Variables, and deploy. No build settings are needed.
Research
Every default is sourced in docs/research:
- finance.md: WACC ranges (Damodaran Jan 2026 CRPs, IRENA, IEA, Lazard), GCC IPP tariffs, preset inputs, tornado conventions
- cases.md: frameworks, difficulty calibration, and the case bank with script-checked arithmetic (the app's case data is generated from this file:
npm run build:cases)
- departments.md: how each function argues in real investment committees, and how leadership teams resolve conflict (RAPID, one-way/two-way doors, pre-mortems)
GCC tariffs and regulations change often; figures marked as estimates are triangulated, not published. Not investment advice.
Project layout
api/ serverless functions: department, brain, coach, health
api/_lib/ Claude client, personas + JSON schemas
js/ app shell, simulation, constellation map, calculator, practice, finance engine
docs/research/ sourced research behind every default
scripts/ case-bank generator
tests/ finance engine tests
MIT licensed. Built with the Claude API.
