▶ Live war-room: https://karnadhar-117722238113.asia-south1.run.app (deployed on Google Cloud Run, Mumbai)
Karnadhar (कर्णधार, Sanskrit): the helmsman who steers the ship through the storm. From a 47-day crisis to a sub-second command.
Built for the ET AI Hackathon 2026 — Phase 2, Problem #2 (AI-Driven Energy Supply Chain Resilience for Import-Dependent Economies), by Mann Sutariya (Pandit Deendayal Energy University).
India imports 88% of its crude; 46% of it transits the Strait of Hormuz — a number this system derives from official customs records, not from a report. Strategic reserves cover ~9.5 days (voluntary — India is outside the IEA's 90-day mandate). When a chokepoint is threatened, economies without response intelligence take 47 days longer to stabilise supply (McKinsey). KARNADHAR compresses signal → executable recommendation to 45 ms.
Every generic "resilience dashboard" treats oil as one number: barrels. Reality: a refinery is an engine tuned to a fuel. Three properties decide whether a barrel is runnable — API gravity (light/heavy), sulphur (sweet/sour), and asphaltene content (the SARA dimension, added on the guidance of Prof. U. K. Bhui, Petroleum Engineering, PDEU). A deep-conversion coker like Jamnagar (Nelson 21.1) digests almost anything but starves on ultra-light condensate; a simple refinery like Mumbai (Nelson 7) cannot touch sour crude.
The headline results (real DGCIS model, 10 refineries × 36 supplier grades):
- Hormuz closure — the naive "oil is oil" plan is infeasible for a grade reason: its blend breaches Visakhapatnam's sulphur ceiling. KARNADHAR's two-stage LP re-sources the full 2,313 kb/d gap feasibly — and prices the hidden cost: the reroute ties up +66 VLCC-equivalents (the ton-mile effect a fungible view never sees).
- Russia sanctioned — naive sends 220 kb/d to refineries that physically cannot run it; KARNADHAR: zero. Usable shortfall 220 → 0.
- Hormuz + Russia (compound) — naive pretends to fill the gap with 546 kb/d of un-runnable crude; counting only runnable barrels, KARNADHAR re-sources 504 kb/d more usable supply and quantifies the honest remaining gap.
- Shadow prices — the LP's duals name the scarcest barrel (heavy Colombian coker feed, ~$4.2k/day per extra kb/d): a procurement priority list that is derived, not opined.
- On the curated stress-demo (
run_demo.py), the wedge is worth $3.15M/day of protected yield (~$1.15bn/yr) — labelled as the demonstration model.
signal → scenario → reroute → brief
GDELT news + glass-box economic grade-aware LP executive memo —
live AIS ships cascade + twin over real DGCIS LLM writes words,
deficit refinery diets never numbers
- Real data spine — official DGCIS port-wise import records (May 2024–Apr 2026) give every refinery its actual crude diet, Hormuz exposure, and realized landed prices. PPAC capacities, public Nelson complexity, crude assay libraries.
- Knowledge graph — the supplier→route→chokepoint→refinery relationship model,
materialised as a typed property graph with per-edge provenance
(
python export_kg.py→ 50 nodes, 163 edges: SHIPS_VIA / SUPPLIES / THREATENS). - General disruption model — block any chokepoint (Hormuz, Bab-el-Mandeb, Suez, Malacca, Cape) and/or sanction any supplier (e.g. Russia = 1,590 kb/d), in any combination.
- Honest engine — under compound shocks (Hormuz + OPEC squeeze) even the optimal reroute leaves demand unmet; KARNADHAR quantifies the gap instead of pretending.
- Twin-deficit cascade — the India-specific vulnerability (reviewed with Lydia Powell, Distinguished Fellow, ORF): a full closure adds ~$191bn/yr in USD outflow, blowing the current-account deficit from 0.7% → 6.1% of GDP. The binding constraint is the balance of payments, not the barrels.
- Multi-commodity lens — the framework generalised to 8 strategic imports (LNG,
pharma APIs, semiconductors, edible oils, fertiliser, coking coal, solar PV) via a
glass-box Import Vulnerability Index; the same disruption scored across every
material (
python run_commodities.py— Hormuz hits 53% of LNG, not just 46% of crude; Malacca is the pharma/electronics artery). - Decision layer, not just a dashboard — every scenario ships with the LP's shadow prices (marginal value of one more kb/d of each scarce grade), the usable-shortfall honesty metric, and the VLCC-equivalent tanker cost of longer voyages.
- Reroute deep-dive — click any disruption and open a full-screen decision page that ranks the alternative sourcing corridors #1..N by the volume the optimizer committed (the LP's own stage-1 priority; cost as tiebreak — not shadow-price-first, which would rank a 3 kb/d scarcity artifact above the 90% lifeline). Each corridor carries a self-defending justification, a BINDING/slack tag, its LP marginal value, and the refineries it feeds; clicking one isolates it on the map with a live derivation panel. A severed-corridors section shows the bigger suppliers that were cut (Iraq > Russia, gone at Hormuz) — proving the rank encodes grade-fit and chokepoint survival, not raw size.
- War-room UI — Next.js + MapLibre command center that draws the optimizer's plan itself: every LP allocation becomes an animated, arc-separated flow line — source → real corridor (Cape where that's the true voyage) → the exact refinery it feeds, width = kb/d, hover for the allocation. True 3D globe / 2D toggle, real AIS vessels, bypass pipelines, a fully interactive knowledge graph (zoom/pan/drag), the multi-commodity lens, a live ticker tape, and scenario switching with the national plan re-solved in tens of milliseconds.
# 1) Engine (Python 3.12)
pip install -r requirements.txt
python run_validate.py # 62/62 automated checks — the proof
python run_real.py # reroute on REAL DGCIS diets, all scenarios
python run_karnadhar.py # end-to-end pipeline: signal → brief in ~45 ms
# 2) War-room (Node 20+)
python export_ui.py # engine → frontend/public/karnadhar.json
cd frontend && npm install && npm run dev # → http://localhost:3000More runnables: run_demo.py (wedge head-to-head), run_cascade.py (economic cascade +
twin deficit), run_signal.py (GDELT 12-day-lead backtest), run_ais.py (live AIS),
run_scenarios.py (scenario library), run_commodities.py (multi-commodity screen),
export_kg.py (knowledge graph), gen_deck_assets.py (deck charts).
karnadhar/
├── engine/
│ ├── refdata.py # DGCIS loader: port→refinery, country→grade/route, Nelson, asphaltene
│ ├── realmodel.py # real refineries (configs DERIVED from diet + Nelson) + general Disruption
│ ├── realopt.py # naive vs grade-aware LP on the real model
│ ├── data.py # curated 15-grade / 7-refinery model (the original wedge demo)
│ ├── optimizer.py # wedge optimizer + honest SPR evaluation
│ ├── cascade.py # glass-box cascade incl. twin-deficit block
│ ├── scenarios.py # scenario library
│ ├── orchestrator.py # timed signal→scenario→reroute→brief chain
│ ├── briefing.py # exec memo (template; Claude drop-in via ANTHROPIC_API_KEY)
│ ├── geo.py # chokepoints, routes, coordinates
│ └── signals/
│ ├── gdelt.py # live GDELT client + cache + labelled fallback
│ ├── agent.py # risk scoring, alert rule, lead-time backtest
│ ├── ais.py # live aisstream.io client (per-chokepoint bboxes)
│ └── extract.py # headline classifier (keyword; Claude drop-in)
├── api/main.py # FastAPI backend
├── frontend/ # Next.js 16 + MapLibre war-room
├── run_*.py # runnable proofs (see above)
├── run_validate.py # 62-check validation suite (exit-code gated; runs in CI)
├── export_ui.py # engine → UI JSON
└── deliverables/ # pitch deck (KARNADHAR_Deck.pptx) + renders
| The brief's "what you may build" | KARNADHAR component |
|---|---|
| Geopolitical Risk Intelligence Agent (news, AIS, sanctions, price signals) | GDELT client + explainable alert rule (12-day backtested lead) · live AIS vessels · sanctions as a first-class disruption axis · realized landed prices from customs records |
| Disruption Scenario Modeller (Hormuz / OPEC+ / Red Sea) | 5-scenario library incl. Red Sea suspension — crude is Cape-insulated (54 kb/d) while edible oils take 18%: commodity-specific arteries |
| Adaptive Procurement Orchestrator (spot pricing, tanker availability, grade compatibility) | two-stage grade-aware LP + shadow-price procurement priorities + VLCC ton-mile cost (+66 tankers on Hormuz) |
| Strategic Reserve Optimisation Agent (drawdown schedules) | engine/spr.py — hold / bridge / ration scheduler with draw rate, depletion %, residual demand-management need |
| Supply Chain Digital Twin (geospatial what-if platform) | 3D-globe war-room + knowledge graph + bypass-pipeline mapping; every scenario re-solved live |
Suggested technologies: agentic multi-agent chain ✓ · geospatial (AIS, pipeline & port mapping) ✓ · predictive analytics & scenario simulation ✓ · knowledge graph (supplier-route-risk-refinery) ✓ · LLM signal extraction (keyword classifier with a key-gated Claude drop-in — words, never numbers) ◐ · RAG over intelligence corpora + sanctions-registry/port-congestion feeds → roadmap (documented, not claimed).
Full source list with links: see SOURCES.md — DGCIS, PPAC, ISPRL,
EIA, IEA, GDELT, aisstream, RBI, MoSPI, crude assays, and the two expert reviewers.
Reproducibility: the raw DGCIS .xls files are government downloads and are
not redistributed here; the repo ships the derived dataset
(engine/data/india_refinery_diets.json, schema-versioned, with the derivation
code in engine/refdata.py). Every run — validation, exports, the war-room —
works from a fresh clone; anyone holding the raw files reproduces the dataset
bit-identically with python -m engine.refdata.
| Element | Status |
|---|---|
| Refinery diets, volumes, landed prices | Real — DGCIS official trade records |
| Refinery capacities | Real — PPAC |
| National import mix, Hormuz exposure | Derived from the real records (46%) |
| Crude assays (API/sulphur/asphaltene) | Public assay values per source grade |
| GDELT signal series | Real (cached June 2025); method identical live |
| AIS vessels | Real live where coverage exists (Malacca verified); Hormuz snapshot labelled — free feed has no Gulf receivers |
| Refinery processing limits | Derived from real diet + Nelson complexity; flagged for expert calibration (in progress with Prof. Bhui) |
- Lydia Powell — Distinguished Fellow, ORF Centre for Resources Management: twin-deficit framing, SPR-as-voluntary-insurance, sourcing-first policy sequence.
- Prof. U. K. Bhui — Petroleum Engineering, PDEU: wedge validated ("That is — you are perfectly right"), SARA/asphaltene dimension.
No API keys are committed (.env is git-ignored). The aisstream key is free; the
engine runs fully without any key (labelled fallbacks).