SupplAI by GIFS

From market signals to smarter supply decisions.

Inspiration

Petrochemical producers live and die by a single question, asked hundreds of times a week: with this much product on hand, whom do we sell it to, in which market, at what price — right now? For PE and PP grades sold across Southeast Asia, that answer shifts daily with volatile resin prices, plant outages, credit exposure, freight, and shelf-life. Today it's made in spreadsheets and gut feel, and a single mediocre allocation quietly leaks margin.

We didn't want another dashboard that describes what happened. We wanted a prescriptive decision co-pilot that senses demand, reads the market, and tells a planner exactly what to do — with the math to back it and a human in control of every commitment.

What it does

SupplAI turns market signals into ranked, explained supply decisions:

• Demand intelligence — a decision command centre that forecasts demand for every product × market, showing where demand is (by country) and what's in demand (by grade) as interactive, cross-filterable charts, each with a confidence band and a self-scored accuracy from walk-forward backtesting.
• Scenario simulator — a planner adjusts assumptions (demand uplift, price/cost moves, safety stock, margin threshold, strategic weights) and runs a real constrained optimiser that computes the profit-maximising allocation under inventory, production, credit, logistics and feasibility limits — then explains each recommendation in plain language.
• Decision workflow — every recommendation can be approved, modified, or rejected with a reason; modifications rescale the economics; everything is audited.
• Market-signal sensing — free-text market news ("force majeure hits Vietnam supply") is turned into structured signals that actually move the plan — raising risk on affected markets or shifting sensed demand on the next run.
• Data import — products, customers, prices and historical sales load from CSV with a validate/commit flow, so the whole platform runs on real numbers.

Crucially, the AI never invents a number. The OR-Tools solver owns every allocation; the AI layer only explains, answers questions, and structures news.

How we built it

• Web: Angular 19 (standalone components, Signals, strict TypeScript) — an enterprise workspace with a visual, chart-first demand command centre. No charting dependency; pure CSS/SVG.
• API: Express 5 + Zod validation + JWT/RBAC, tenant-aware, with a repository pattern that runs on PostgreSQL via Prisma or in-memory for a zero-setup demo.
• Optimiser: an isolated FastAPI + Google OR-Tools (CBC) service that solves the constrained allocation and returns diagnostics (binding constraints, unmet demand, excluded opportunities). A deterministic in-process heuristic transparently takes over if the solver is unreachable — and never falsely claims OPTIMAL.
• AI layer: a provider contract with a deterministic, key-free implementation as first-class default, and a live Claude (claude-opus-4-8) provider for richer explanations, plan Q&A, and news→signal extraction. The platform is fully functional with no paid AI dependency.
• Data: a reproducible 8-grade × 5-market SE-Asia dataset, plus a grounded 24-month sales / 18-month price pack for realistic demos.

Challenges we ran into

• Getting the optimiser's objective right. An early version mixed a margin ratio with dollar risk costs, so raising a price paradoxically dropped the allocation. We reworked the objective to a consistent dollar unit_margin so the local heuristic mirrors OR-Tools exactly.
• Making news genuinely change the plan. Wiring extracted signals through persistence into the optimiser's demand and risk inputs — and proving a supply-disruption headline actually lowers plan margin — took careful classification (a "disruption" that reads neutral still has to raise risk).
• A missing migration. Our initial Prisma migration predated the sales table, so the Postgres path would have failed at runtime — caught it by diffing schema vs. migrations and adding the missing one.
• Keeping AI honest. Enforcing that the AI boundary receives only authorised, already-computed facts and can never mutate an allocation — with a deterministic fallback so a failed API call degrades gracefully.
• Toolchain gremlins: duplicate rxjs breaking Angular operator inference, a Docker build context mismatch, and a CSV column shifted by an unquoted comma in source data.

Accomplishments that we're proud of

• A runnable, end-to-end vertical slice — login → forecast → optimise → explain → decide → audit — verified in a real browser and against live Postgres + OR-Tools.
• The AI boundary discipline: prescriptive and trustworthy, because the solver owns the numbers and the platform works with zero API keys.
• Demand sensing that closes the loop: importing real history lifted a forecast from 558 → 768 t and pushed board-wide backtest accuracy to ~96%; a committed news signal measurably shifted the plan.
• A decision command centre that a non-technical planner can actually read — and the data lands where reality says it should (Vietnam, the world's #1 PP importer, ranks as the peak market).
• 69 passing API tests, a clean production build, and the whole thing spins up with one docker compose up.

What we learned

• Prescription beats description — but only with a human in the loop. The value isn't the number; it's the ranked, explained, overridable recommendation with an audit trail.
• Draw a hard line around the LLM. Letting AI explain and structure while a solver computes gives you both credibility and delight. Blur that line and you lose trust.
• Deterministic-first is a feature. A key-free default means the demo always works, tests are reproducible, and cost is never a blocker to adoption.
• Real trade data validated our instincts — the public flows (Vietnam #1 PP importer; Indonesia sourcing from Thailand/Singapore/Malaysia) matched the market structure we'd modelled.

What's next for SupplAI

• Live data feeds: wire UN Comtrade / WITS demand volumes and ICIS/businessanalytiq price series directly, replacing sample data.
• Automated e2e + CI: commit the Playwright flows as a suite with lint/typecheck/test on every push.
• Financial precision: migrate money/quantity columns to Decimal.
• Richer signal sensing: continuous news ingestion with confidence-weighted, decaying signals feeding a rolling risk model.
• More decision surface: inventory and production-plan optimisation, multi-period planning, and OpenAPI docs for the Express API.
• Pilot with real ERP data to quantify the margin uplift against today's spreadsheet workflow.

Built With

• angular.js
• claude
• codex
• css
• express.js
• fastapi
• postgresql
• prisma