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]
This is the beginning of the README. Read the rest at the original. Read the full README ↗
