gabrielwang.ai
Applied AI · Investment & Trading Operations
Real systems, rebuilt to share

AI you can click,
not AI you read about.

I build AI for the work between the trade and the audited number — in funds and in physical commodities. Daily NAV and mark-to-market, reconciliation and month-end close, counterparty credit and the receivables ledger, documentary compliance on cargoes, proprietary research and investment due diligence, market intelligence that reads short- and long-term relative strength instead of just fetching prices. These systems run unattended, in production, every day. What's here are working rebuilds — cut down several grades, onto synthetic data and fictional names, so they can be shared.

See what's in the workshop → How this works

The idea

Most AI content tells you what's possible. This shows you — on the workflows you sign off on.

Each edition takes one real workflow from fund or physical-trading operations and ships it as a demo you can poke at. No signup wall, no pitch.

01

One real workflow

NAV and mark-to-market, position and cash reconciliation, LC document checks on physical cargoes, receivables and payables against counterparty limits, month-end close, due-diligence screens, market briefings — the work around every trade, from first look to audited number.

02

A demo you can click

Interactive, in your browser, on synthetic data. Step through the pipeline, break something, watch how the system responds.

03

Guardrails over magic

The interesting part isn't that AI drafts the number — it's the integrity gates that stop a bad number from ever reaching stakeholders.

04

The CFO's read

What it costs to build, where it breaks, what it does to control, risk and the cost of running the function — and whether it's worth doing. If it wasn't worth building for myself, it isn't here.

A workshop, not a storefront

Nothing here is for sale — no pricing, no booking link, no funnel.

I built these systems because I needed them, and they've paid for themselves many times over — in hours, and in errors that never happened. Sharing the working patterns costs me little and might save you a quarter of trial and error. That trade seemed obviously worth making.

What you see is deliberately de-tuned.

The originals run in production with real money, real counterparties and real consequences. Out of respect for privacy and confidentiality, every demo is rebuilt several grades down before it's published: synthetic data, fictional names, generic workflows, and a fraction of the moving parts. What a five-minute click-through can't carry — live data feeds, multi-entity books, the edge cases and watchdogs accumulated through years of production failures — is exactly what the originals are made of. The plumbing, the guardrails and the judgment calls shown here are the real thing. The numbers, and the difficulty, are not.

In the workshop

Ten interactive demos, one synthetic world — a fund and a physical commodities book. Start anywhere — they cross-link.

001

The principal's relationship memory. Live demo

30 seconds of WhatsApp voice after each meeting; the bot files people, threads, promises and personal anchors. Months later: trip briefs that surface your own overdue promise before you land, a Thursday promise sweep, and quiet-relationship flags tuned to each contact's own rhythm.

Launch the demo →
002

Daily NAV & reconciliation — with integrity gates. Live demo

Step through a fund's daily NAV build on synthetic broker data. Then inject a failure — a wrong price, a missing trade, an odd-lot — and watch reconciliation catch it and hold distribution. The system that refuses to send bad numbers.

Launch the demo →
003

Your AI can read the fund — not touch it. Live demo

An assistant wired to Edition 001's fund through a read-only MCP server: seven tools that answer anything — NAV, P&L drivers, weights, recon — and zero that can write, book, mark or send. Ask it to trade, and watch the wire refuse.

Launch the demo →
004

Shock the book. Live demo

Stress Edition 001's fund on purpose: the FX move that hits one sleeve, Nvidia −30%, the coordinated bad week — then a 20% redemption priced at post-shock NAV. Honest arithmetic, renormalised weights, and the cash ladder. No risk theatre.

Launch the demo →
005

The million-dollar paper trail. Live demo

A US$14.8M iron-ore cargo is paid on documents, not ore. Eight strict-compliance checks — exact strings, dates, wet-to-dry tonnage maths — five ways to fail them, and the cure memo for each, before the bank finds it first.

Launch the demo →
006

The month-end pack that writes itself. Live demo

One click closes the month: grid column, compounded YTD, attribution that foots exactly, signed capital flows. Then break it — run it twice (idempotent), fail the recon (pack HELD), drop funding mid-month (NAV moves, P&L doesn't).

Launch the demo →
007

Dormant counterparty reactivation — the desk sheet that ranks who to call. Live demo

A trading desk spots counterparties who've gone quiet against their own buying rhythm — not a blunt 90-day rule — ranks them by revenue at risk, and drafts the outreach.

Launch the demo →
008

LLMs judge, code counts. Live demo

Three judge personas score four listing candidates in words, with reasons. Code does every number: robust medians, weighted lenses that flip the ranking live, and a disagreement queue routed to humans — the split that makes AI scoring auditable.

Launch the demo →
009

Automation that can't fail silently. Live demo

The ops doctrine behind unattended systems: three-outcome probes, page only on confirmed-down, quiet hours that defer but never drop, delivery gates that escalate before the deadline. Break the fleet six ways and watch what reaches the phone — and what deliberately doesn't.

Launch the demo →
010

The agent perimeter. Live demo

Fire six messages at a 24/7 WhatsApp AI — a stranger, a prompt injection, a request for its keys, a phone swap. Five layers decide what it answers, what it confirms first, and what it never even sees. The empty pane is the product.

Launch the demo →

Who's behind this