Type a product. Pick source countries. Choose a transport mode. Get a ranked, cited, ML-scored comparison of real manufacturers — in under 60 seconds.
- Why GreenChain
- What it does
- Product demo
- Architecture
- User journey
- The ML pipeline
- Data sources
- Tech stack
- Quickstart
- API reference
- Market & opportunity
- Roadmap
- Team
- License
Global supply chains are responsible for over 60% of corporate greenhouse gas emissions — yet the vast majority of procurement decisions are made with zero visibility into those emissions. Existing ESG software costs upwards of $50,000/year, requires months of onboarding, and is built for compliance teams — not for the fast, real-world decisions that procurement managers make every week.
The landing page — GreenChain frames the decision before a single PO is cut.
Procurement managers, ESG leads, and climate-aware founders face the same three problems:
| Problem | Today's reality | GreenChain's answer |
|---|---|---|
| No common yardstick. | Emissions disclosures are inconsistent, self-reported, and often missing. | We normalize across USEEIO, Ember, GLEC, and ND-GAIN to a single composite score. |
| Manual research. | Analysts spend days reading sustainability PDFs manufacturer by manufacturer. | A Dedalus agent swarm does it in parallel, in under a minute. |
| No counterfactuals. | "What if we switched to rail?" requires a new spreadsheet. | Flip the transport toggle — the ranking re-computes instantly, client-side. |
One sentence: You describe what you're sourcing — product, destination market, quantity, candidate countries, and transport mode — and GreenChain automatically discovers real manufacturers for the components, scores them across five environmental dimensions, and ranks them on a live dashboard.
The live dashboard: supply-chain graph on the left, per-supplier drill-down in the middle, animated geographic routes on the right.
- 🌐 Interactive globe view — manufacturers pinned on a 3D globe with live-animated transport arcs.
- 🕸️ Force-directed supply chain graph — see the product → country → manufacturer topology update as agents stream in results.
- 🤖 Agent swarm discovery — Dedalus orchestrates parallel
web_search,fetch_url, and custom Python tools for every country in the query. - 📊 XGBoost quantile scoring — q10 / q50 / q90 emissions intervals trained on 1,016 NAICS industries × 179 countries.
- ⚖️ Adjustable weight sliders — users re-weight the 5 scoring dimensions and see rankings shift live.
- 🚢 Instant transport rescore — flipping sea → air runs the GLEC formula in the browser, no network round-trip.
- 📄 Downloadable memo — a final agent writes a citable procurement recommendation.
Screenshots below are captured from the production build. To explore interactively, follow the Quickstart.
Drop a scenario CSV (one row per component) or click Continue with demo to load a pre-seeded 20-node graph over 16 routes.
While the agent swarm runs, a globe pre-renders the route topology — the UI never shows a spinner in a vacuum.
The dashboard is a resizable three-pane workspace:
| Pane | What lives there |
|---|---|
| 🕸️ Supply Chain Graph (left) | Force-directed graph: root = product, mid-nodes = components, leaves = manufacturers. Edges color by env-rating; nodes grow with total tCO₂e as scoring streams in. |
| 🔍 Supplier drill-down (center) | Click any node for a radial score breakdown, certifications, disclosure status, and alternate suppliers for the same component. |
| 🌐 Geographic Routes (right) | 3D globe (Three.js) with animated transport arcs whose color encodes transport mode and whose travel speed matches the chosen mode. |
| 💬 Prompt bar (bottom) | "remove suppliers who don't like cats or have bad mojo" — natural-language filter over the current result set, routed to a second-pass scenario editor. |
Key files:
components/dashboard/interactive-globe.tsx,components/dashboard/supply-chain-graph.tsx,components/dashboard/prompt-bar.tsx.
GreenChain is a thin FastAPI orchestrator in front of a Dedalus agent swarm, with all ML scoring and data lookups happening in-process post-swarm.
┌─────────────────────────────────────────────────────────────────────┐
│ NEXT.JS FRONTEND │
│ Globe (Three.js) Supply-chain graph (D3) Prompt bar + results │
└─────────────────────────────┬───────────────────────────────────────┘
│ POST /search (product, countries, mode)
▼
┌─────────────────────────────────────────────────────────────────────┐
│ FASTAPI BACKEND │
│ (pass-through orchestrator) │
└─────────────────────────────┬───────────────────────────────────────┘
│ runner.run(input, model, mcp_servers, tools)
▼
┌─────────────────────────────────────────────────────────────────────┐
│ DEDALUS AGENT SWARM │
│ │
│ 🔎 Discovery agent ─► brave-search-mcp │
│ 📜 Certification agent ─► fetch_url (optional Dedalus Machine) │
│ 🧮 Tool-using agent ─► lookup_emission_factor │
│ calculate_transport_emissions │
│ score_certifications │
│ ✍️ Memo agent ─► writes procurement recommendation │
└─────────────────────────────┬───────────────────────────────────────┘
│ raw JSON: manufacturers + certs + URLs
▼
┌─────────────────────────────────────────────────────────────────────┐
│ ML SCORING PIPELINE (in FastAPI) │
│ │
│ XGBoost quantile regression (q10/q50/q90) — manufacturing tCO₂e │
│ GLEC framework × port distance ─ transport tCO₂e │
│ Ember grid carbon ─ local electricity intensity │
│ ND-GAIN climate risk ─ physical-risk adjustment │
│ Certification multiplier + composite 0–100 normalizer │
└─────────────────────────────┬───────────────────────────────────────┘
│ SSE stream of scored manufacturers
▼
Back to frontend
- Dedalus does all agent orchestration. No hand-rolled tool-use loops, no asyncio semaphores. One
runner.run()call per search. - ML is in-process, not agentic. Scoring is deterministic and reproducible; judges and investors can audit the formula.
- Transport toggle never re-hits the backend. The GLEC factor table lives in the browser — flipping sea → air is sub-frame.
- Optional Dedalus Machine isolation. Set
GREENCHAIN_USE_DEDALUS_MACHINE=1to run the agent's URL fetches inside a KVM-isolated Linux VM (provisioned at startup, deleted on shutdown). Falls back to localhttpxon any error.
- Land → cinematic hero page frames the problem.
- Launch → onboarding overlay collects product, countries, destination, transport mode.
- Search → agents stream results into the globe, graph, and result drawer in real time.
- Compare → flip transport modes, adjust weight sliders, filter to only certified suppliers.
- Decide → download the auto-generated recommendation memo with cited sources.
| Dimension | Method | Source |
|---|---|---|
| Manufacturing emissions | XGBoost quantile regression (q10 / q50 / q90), features: NAICS code × ISO country × grid carbon. | USEEIO v1.3 (1,016 NAICS) + Ember + CDP supply-chain disclosures. |
| Transport emissions | GLEC factor (kgCO₂e/tonne-km) × shipment weight × port-to-port distance lookup. | Smart Freight Centre GLEC v2.0 + SeaRates port matrix. |
| Grid carbon | Per-country electricity carbon intensity (gCO₂/kWh). | Ember Climate (179 countries). |
| Certifications | Weighted multiplier across ISO 14001, CDP A/B/C, SBTi committed/achieved, B Corp. | Parsed from manufacturer sustainability pages by agent. |
| Climate risk | Physical risk (flood, heat stress) per-country. | Notre Dame GAIN (167 countries). |
The composite is a weighted 0–100 score where each dimension's weight is user-adjustable in the UI. Default weights:
DEFAULT_WEIGHTS = {
"manufacturing_emissions": 0.40,
"transport_emissions": 0.25,
"grid_carbon": 0.20,
"certifications": 0.10,
"climate_risk": 0.05,
}Models are vendored under
backend/ml_runtime/models/(~9 MB) and loaded once at FastAPI startup. Training code lives inbackend/ml_runtime/ml/.
Every number in GreenChain traces back to a free, citable, public dataset — so when an investor, auditor, or judge asks "how do you know this?" we have an answer.
| Dataset | Scope | Link |
|---|---|---|
| EPA USEEIO v1.3 | 1,016 NAICS industries × emission factors (tCO₂e / $1M) | epa.gov/useeio |
| Ember Climate | 179-country electricity grid carbon intensity | ember-climate.org |
| GLEC Framework v2.0 | Global Logistics Emissions Council transport factors | smartfreightcentre.org |
| ND-GAIN Country Index | 167-country climate vulnerability & readiness | gain.nd.edu |
| CDP Supply Chain | Corporate emission disclosures (training data) | cdp.net |
Frontend
- ⚛️ Next.js 16 (App Router, Turbopack dev / Webpack build)
- 🎨 Tailwind v4 + shadcn/ui + Radix UI primitives
- 🌐 Three.js interactive globe
- 🕸️ D3-style force-directed graph (custom, no d3 dep)
- 📦 React 19, React Server Components, SSE streaming
Backend
- 🐍 FastAPI (async) + Uvicorn
- 🤖 Dedalus Labs
AsyncDedalus/DedalusRunner— agent orchestration - 🔎 Brave Search MCP for manufacturer discovery
- 🧠 XGBoost quantile regression + scikit-learn
- 🗄️ SQLite for reference tables (USEEIO, Ember, GLEC, port distances, ND-GAIN)
- 🔒 Optional Dedalus Machines (KVM-isolated VM) for sandboxed
fetch_url
Models
anthropic/claude-sonnet-4-6via Dedalus passthrough (discovery + certification agents)- K2 Think V2 — two-pass scenario editor (parse-safe JSON mutations)
- Gemini — used in the memo-generation pipeline
- XGBoost q10/q50/q90 emissions model (vendored)
- Python 3.11+
- Node.js 20+
- API keys for Dedalus, Anthropic, and Brave Search (free tiers cover the demo)
cd backend
cp .env.example .env # fill in DEDALUS_API_KEY, ANTHROPIC_API_KEY, BRAVE_API_KEY
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt # installs dedalus-labs, fastapi, xgboost, …
cd .. # IMPORTANT: run uvicorn from repo root
uvicorn backend.main:app --reload --port 8000Open http://localhost:8000/docs for the interactive Swagger UI.
Troubleshooting
No module named 'dedalus_labs': the process is using a Python different from where you ranpip install. Activatebackend/.venvin the shell you runuvicornfrom (or point your IDE interpreter to it).
cd frontend
npm install
npm run dev # http://localhost:3000The frontend expects the backend at http://localhost:8000.
| Key | Where to get it | Required? |
|---|---|---|
DEDALUS_API_KEY |
dedaluslabs.ai | ✅ |
ANTHROPIC_API_KEY |
console.anthropic.com | ✅ |
BRAVE_API_KEY |
api.search.brave.com (free: 2k q/mo) | ✅ |
GREENCHAIN_USE_DEDALUS_MACHINE |
Set to 1 to sandbox fetch_url in a Dedalus Machine VM |
❌ |
GREENCHAIN_ALLOW_MOCK_COMPONENT_SEARCH |
Set to 1 to fall back to mock manufacturers when keys are missing (demo-only) |
❌ |
If /search returns 502, read the detail message — it's almost always a missing key or an upstream Dedalus / network failure.
| Method | Path | Purpose |
|---|---|---|
GET |
/health |
Liveness probe. |
POST |
/search |
Run the Dedalus swarm + ML scoring; returns a ranked list. |
POST |
/score |
Score pre-collected candidates (skips the agent call). |
POST |
/rescore-transport |
Recompute transport emissions under a different mode. |
curl -X POST http://localhost:8000/search \
-H "Content-Type: application/json" \
-d '{
"product": "cotton t-shirts",
"quantity": 10000,
"destination": "US",
"countries": ["CN", "PT", "BD"],
"transport_mode": "sea",
"target_count": 9
}'Two search modes:
- Per-country — pass a non-empty
countriesarray; the agent finds manufacturers in each. - Global — pass
countries: []; the agent picks countries itself, biased toward geographic diversity.
| # | Product | Countries | Highlight |
|---|---|---|---|
| 1 | Cotton t-shirts (10k units) | CN, PT, BD | Portugal ranks #1 on sea. Flip to air → China transport emissions jump ~55×. |
| 2 | Circuit boards (5k) | TW, VN, DE | Filter to iso14001 only — watch results prune live. |
| 3 | Automotive components (20k) | DE, MX, IN | Road vs. rail toggle changes the ranking on short-haul routes. |
- 📈 ESG software is a $1.4B → $4.3B market by 2027 (Verdantix, 2024).
- 🏭 Scope 3 reporting is mandatory under the EU CSRD (~50k companies in scope from FY2024) and California SB 253.
- 📦 Procurement teams make ~$13T/year in global sourcing decisions (McKinsey, 2023) — and none of the incumbents (EcoVadis, CDP, Watershed) give a live, counterfactual-ready comparison.
| GreenChain | EcoVadis | Watershed | Manual spreadsheet | |
|---|---|---|---|---|
| Time to first ranked list | ~45s | Weeks | Days | Days–weeks |
| Counterfactual transport mode toggle | ✅ instant | ❌ | ❌ | ❌ |
| Cited public data sources | ✅ 5 datasets | Partial (private scores) | Partial | Varies |
| Agent-driven auto-discovery | ✅ Dedalus swarm | ❌ (seller-submitted) | ❌ | ❌ |
| Works without supplier onboarding | ✅ | ❌ | ❌ | ✅ |
- Procurement & sourcing managers at mid-market brands.
- ESG / sustainability leads under CSRD or SB 253 reporting pressure.
- Climate-aware founders making early sourcing decisions that lock in emissions for years.
- Private-market investors doing ESG due diligence on portfolio companies.
- v0.1 — HackPrinceton 2026 — live agent swarm, XGBoost scoring, interactive globe, transport toggle.
- v0.2 — Cost integration. Layer in unit cost and lead time so procurement teams can see the Pareto frontier of environmental vs. financial performance.
- v0.3 — Richer certification signals. Beyond ISO 14001 / CDP / SBTi — parse TCFD, CSRD disclosures, and SBTi target vintages.
- v0.4 — Deeper Scope 3 upstream tracing. Agents follow bill-of-materials one hop deeper into Tier-2 suppliers.
- v0.5 — Real-time grid carbon. Swap annual Ember averages for live marginal-intensity APIs (Electricity Maps, WattTime).
- v0.6 — CSRD / SB 253 export. Map results to required disclosure schemas for audit-ready output.
- v1.0 — Enterprise pilot with 3 design-partner procurement teams; SSO + audit log; hosted Dedalus Machines per tenant.
Built in 36 hours at HackPrinceton Spring 2026 for the Dedalus track "Best agent swarm hosted on Dedalus Containers." Full project write-up on Devpost.
"Agent reliability is an engineering problem, not a prompting problem." The hardest single problem was signal quality — Brave Search returns a mix of real manufacturers, directory listings, trade aggregators, and SEO-spam factories, so filtering and disambiguation had to be baked into the swarm itself. Working with fragmented environmental datasets (USEEIO, Ember, GLEC, ND-GAIN, CDP) in one pipeline taught us how much normalization work sits between "public data exists" and "you can act on it."
We shipped a complete, end-to-end product. The XGBoost model generalizes meaningfully, the Dedalus architecture came out exactly as designed, and the K2 Think V2 two-pass scenario editor produces parse-safe, internally-consistent JSON reliably.
Interested in investing, piloting, or contributing? Open an issue or reach out via the contact on the team's GitHub profiles.
MIT — see LICENSE if present; otherwise treat as MIT for the hackathon submission.
Made with ☕ at HackPrinceton Spring 2026 · Powered by Dedalus Labs + Anthropic Claude