An AI macro-telemetry early-warning system for global food security — thirty to ninety days of warning before a chokepoint closes, and a local buyer for the cargo it strands when one closes anyway. Built for the governments of import-dependent nations, not for the corporations that already have this intelligence.
All the grain, four lanes wide
One closure, four weeks later
Chokepoint Sentinel is a public-good intelligence layer, not a commercial logistics platform. It watches the fourteen physical chokepoints that the world’s grain actually moves through, simulates what a restriction at any one of them does to national food reserves downstream, and hands that answer to the governments who have no other way to get it.
And when the warning comes too late, it places the food rather than watching it rot. A closure does not make a cargo disappear; it lands the ship somewhere it was never bound for. Grain destined for Egypt is put ashore at Djibouti and becomes stranded — physically present, commercially unsellable, on a spoilage clock. Sentinel already knows the vessel, the commodity, the tonnage and the port it was pushed to, so the same telemetry that raised the alarm is used to match that cargo to a pre-vetted regional buyer within days. That is Layer 5, and it is the difference between an alarm and an outcome.
The reason it is built top-down, from satellites and public ship-tracking rather than from corporate systems, is the single hardest constraint in this field: most enterprise supply-chain data is not usable by AI, and no competitor will put its proprietary data on a rival’s platform. Waiting for that to change is waiting for decades. So we do not ask.
Why we are building it: when the Mississippi floods, a US farmer receives a government bridge payment. When the Black Sea closes, an import-dependent nation has no such buffer — and finds out too late to act. The intelligence that multinationals already buy is exactly what these governments cannot afford. That asymmetry is the problem this exists to close.
Schematic, not a survey map. Phase 1 covers three chokepoints; the remaining eleven arrive in Phase 4.
The global supply chain crisis is not a single failure. It is a set of interconnected weaknesses, and they fall hardest on Low-Income Food-Deficit Countries. The first five are the published diagnosis. The sixth is what the first five leave behind after the disruption has already happened, and it is the one most proposals never reach.
615 million tonnes of grain move through just 14 systemically important chokepoints each year. All but one have suffered interruptions lasting days to months in the last fifteen years. Djibouti depends on Black Sea wheat; Honduras on US Midwest maize down the Mississippi.
Cited
Disruptions are no longer black swans; they are chronic. Trade restrictions and climate events cascade into input-cost inflation, and historically into sovereign debt stress and unrest — the 2010–11 Arab Spring followed a Black Sea wheat spike.
Cited
Supply chain operations are defined by exceptions — delays, stockouts, weather. The binding constraint is decision speed: a human team takes days to model what a single port closure does three links downstream.
Cited
74% of enterprise supply-chain data is not AI-ready, held in silos and legacy systems. Any plan that begins “once global corporations clean their data” is a plan measured in decades.
Cited
49% of organisations piloted generative AI in supply chains in 2024. 4% scaled it. Enterprise pilots die because they require competing shipping lines and ports to pool proprietary data on one platform. They will not.
Cited
A closure does not make a cargo disappear. It lands the ship somewhere it was never bound for: wheat contracted to Egypt is put ashore at Djibouti and becomes stranded — physically present, commercially unsellable, accruing storage and demurrage at a port it was never booked into while its own shelf life runs down.
Finding a regional buyer for a bulk food cargo is still a fortnight of telephone calls made by people who do not know the cargo exists. The seller writes off millions; the food rots inside a country that is short of it. Warning prevents the diversions it can. This is the cost of the ones it cannot.
Our framing
Each layer names the obvious approach it rejects, and why. That is the part of an architecture worth reading. The first four detect, simulate, protect and deliver — everything that happens before a disruption. The fifth is what happens after one, and it is the layer that turns a warning into a moved cargo.
We aggregate data that already exists. Computer vision over commercial satellite imagery — counting anchored vessels, measuring canal water levels, detecting port congestion — fused with open AIS ship tracking, macro-weather feeds and aggregated third-party telemetry.
This is how the 74% problem gets bypassed rather than solved: we build a clean top-down layer instead of waiting for a dirty bottom-up one.
What AI is genuinely good at is simulating cascading impact from current variables. When a risk threshold is breached — drought in Panama, say — the twin runs ten thousand permutations of global trade flow and answers a specific question: if the Panama Canal restricts draft by 20%, how many days until maize reserves in Honduras fall below thirty?
Decision time moves from weeks to hours. That is the whole claim.
Competitors will not share data on a public ledger. So we use federated learning and zero-knowledge proofs: ports, shipping lines and governments improve the shared model without exposing commercial data to one another.
This is the antitrust and data-sovereignty problem that kills supply-chain platforms, answered at the architecture level rather than in a terms-of-service document.
The wedge, and the reason the rest exists. Early-warning dashboards built for the governments of import-dependent nations and for agencies such as the World Food Programme:
— 30, 60 and 90-day food security risk alerts.
— Policy simulation: model the effect of a 15% Black Sea export tariff on our national bread price.
— Pre-calculated alternative routing for state import entities.
This gives vulnerable nations the predictive intelligence that multinational corporations already hold and do not share.
The first four layers assume the warning arrives in time. Sometimes it does not, and the vessel is turned into a port nobody planned for. At that moment Sentinel holds something no one else does: the vessel, the commodity, the tonnage, the shelf life and the berth it was actually pushed to.
Layer 5 spends that knowledge on the only question left — who, within reach of this port, can take this cargo this week. It is an automated matching engine between stranded bulk commodities and pre-vetted regional buyers, and it is deliberately the narrowest commercial surface on this page.
Layer 5, set out mechanically. When a route is disrupted and a ship is forced into a port it was never bound for, the cargo is stranded and a clock starts. This is how it gets placed — automatically, locally, and without a single telephone call.
This layer now has a portal of its own — the Stranded Cargo Liquidity Hub →
Swipe the schematic sideways →
Schematic of the sequence, not a screenshot — none of it is deployed. The bracket is the claim: every step right of the diversion is automated, and the elapsed time between a cargo registering at an unplanned berth and a signed local offer is the number this layer is judged on.
When the twin detects or predicts a route failure, it ranks the nearest safe, operational ports by draft, berth availability, storage capacity and inland reach, rather than by distance alone. The cargo is diverted to one of them. The registration of that arrival is the event every step below hangs off — the moment a cargo becomes stranded is the moment the system knows it is stranded.
The instant the cargo registers at the alternate port, Sentinel scans a pre-vetted regional buyer registry — flour mills, national and provincial food reserves, nearby commercial traders — and returns only those able to take that commodity, in that tonnage, inside the remaining shelf life. Vetting happens in advance, with the host government. Nobody is discovering a counterparty during a crisis.
The cargo is now in a market it was never priced for. The system computes a distress discount against the prevailing local rate from a formula published in advance — remaining shelf life, storage and demurrage at that berth, tonnage, and haulage to the buyer’s silo — and shows both sides the working. Roughly ten per cent below local spot is the illustrative shape of it. The number is disclosed, symmetric and auditable.
Matched buyers receive a signed notification: 5,000 tonnes of milling wheat at Djibouti, 10% below local spot, twelve days of shelf life remaining. Acceptance transfers title against a digital contract, the seller is compensated on that instrument instead of waiting out a claim, and the grain moves into local silos while it is still food rather than waste.
This is the only layer that sits near money changing hands, so its boundaries are part of the design rather than a policy written afterwards. A matching engine for food in a shortage can go wrong in obvious ways, and each of these exists to close one of them.
We never take title, and we never hold cargo. Sentinel is the matching and pricing layer between a seller and a vetted buyer. It does not trade the commodity, does not warehouse it, and has no inventory position that a deeper discount would benefit.
The fee is fixed, and it is not a share of the discount. A percentage of the spread would pay us to price the seller down in a famine. A flat matching fee per completed transfer cannot.
No bidding, no auction, no public order book. Buyers see an offer computed from the published formula, not each other’s hands. Distressed food cargo is not a market that should be watchable in real time by anyone with a screen.
Standing: the registry is vetted with the host government, and a national or provincial food reserve holds right of first refusal on any cargo landed in its territory before a commercial trader is shown the offer. The engine is a placement tool for the country the cargo landed in, not a channel for taking that cargo out of it.
Why this is worth building: it stops physical food waste, it compensates the original seller in days rather than at the end of a claim cycle, and it feeds the local market that the diversion happened to put the food in front of — without a single telephone call or manual search. It also completes the argument on the rest of this page. An early-warning system that is right and early enough saves the shipment; one that is right and late still has to answer for the cargo sitting on a quay. This is that answer.
Most proposals list what they include. The useful list is the other one — because every item here is an idea that sounds good in a pitch and fails in deployment, and knowing that is the difference between a demo and a system.
TradeLens — the largest attempt at this — shut down. Competing carriers will not write their commercial position to a shared ledger. We use zero-knowledge proofs and federated learning instead, which give the same integrity guarantee without asking rivals to trust each other.
Instrumenting millions of containers is a capital programme, not a software project, and it puts hardware procurement on the critical path of a famine warning. We read the satellites and AIS feeds that already exist.
It does not. We simulate cascading consequences of conditions that have already been observed. That is a narrower claim and it is one we can actually defend under questioning.
49% piloted, 4% scaled. We are not building a tool that dies in a six-month enterprise pilot; we are building a sovereign public-good dashboard that creates bottom-up pressure for the industry to adopt.
The obvious version of Layer 5, and the wrong one. A public order book for distressed food cargo is a price-discovery machine pointed at a shortage: it invites open speculation on a hungry country’s deficit, it clears nothing until it has deep liquidity on both sides, and running it makes us a licensed exchange operator rather than an intelligence layer. Replaced by matching against a government-vetted local registry at a published formula — no auction, no order book, no counterparty watching another’s hand.
Two kinds of number appear below and they are not the same kind of thing. Figures marked Cited are published statistics from outside this project. Figures marked Projected are ours, and each one shows the assumption it rests on.
Government bridge payments in a single year, in the United States alone, to offset temporary trade-market disruption. Apply that multiplier to a country with no fiscal buffer.
Grain moving annually through the fourteen critical chokepoints. A 10% disruption is 61.5 million tonnes of food that does not arrive.
Logistics technology deals signed at the 2024 Saudi logistics forum — evidence that sovereign appetite for supply-chain resilience already exists and is funded.
Resilience has always been treated as a cost centre. These five modules are what turn it into something self-sustaining — commercial and sovereign efficiency gains paying for the dashboard that LIFDC governments use free. The first four act before a disruption; the fifth recovers value after one that got through.
WFP emergency-premium saver. When a chokepoint fails, commercial freight rates spike 300–500% overnight and the WFP pays it or flies the cargo. Algorithmic pre-positioning moves strategic reserves before the spike.
Assumption: emergency sea and air freight costs 5–10× standard rates, and Sentinel enables 10% of at-risk cargo to be pre-positioned 30 days ahead of a predicted closure.
Transshipment arbitrage. On a >70% probability of Red Sea or Suez disruption, the model prices diverting to Jeddah or King Abdullah Port and moving cargo overland by rail to the Gulf.
Assumption: routing a large bulk carrier around Africa adds 10–14 days and roughly $1M in fuel and time; a land bridge cuts that to about 3 days. Designed to integrate with emerging infrastructure such as the Saudi Landbridge — not a commitment from any operator.
Parametric insurance oracle. We do not underwrite and we do not pay claims. We are the verified data feed a parametric policy triggers on, so a settlement that takes 6–12 months takes a week.
Assumption: parametric policies settle on a verified index rather than a loss adjustment. A LIFDC receiving funds in seven days can buy alternative supply before the price moves. Revenue is a fraction of a verification fee, not a share of the payout.
Green corridor carbon value. Avoiding a congested port burns less fuel. That avoided emission is currently unmeasured and unmonetised; we verify it into a form the voluntary carbon registries accept.
Assumption: global shipping emits roughly 1 Gt CO₂ a year; optimising 1% of grain shipping saves on the order of 10 Mt, valued conservatively at $15/tonne. This revenue is earmarked to subsidise the LIFDC dashboards.
Stranded-cargo matching. A diverted bulk cargo with no local buyer loses on two clocks at once — storage and demurrage at a berth it was never booked into, and the commodity’s own shelf life. Matching it locally at a disclosed discount returns most of the cargo’s value to the seller in days, and puts the food into the market that is short of it.
Assumption: a distress discount of about 10% below the prevailing local rate is illustrative, not a fixed rate; the real figure comes out of the published formula. The comparison is not against a full-price sale — it is against a multi-week stranding in which storage, demurrage and spoilage take a far larger share and, past the shelf life, all of it. Our revenue is a fixed matching fee per completed transfer, never a share of the discount.
The resilience dividend. The loop is the argument: commercial actors save money — on pre-positioning, on rerouting, on cargo placed instead of written off — the savings are captured as fixed fees and verified carbon value, and that pays for the nations who cannot.
On sourcing: every figure marked Cited is drawn from published research and industry reporting compiled in this proposal’s research file, and the full source list travels with the submission document. Figures marked Projected are modelled by us from the assumptions printed beside them; none of them describes money already saved.
No hardware on the critical path, and the trust layer deliberately after the pilot — federated learning is worth building once there is a model worth contributing to. Layer 5 rides the same order: the buyer registry is human work done with governments in Phase 3, and only then is there anything to match.
Satellite computer-vision and AIS aggregation pipelines for three priority chokepoints: the Black Sea, the Panama Canal and Bab al-Mandab. No hardware deployment.
Train the digital twin on historical disruption data from 2010–2024. Pilot the sovereign dashboard with two or three LIFDC governments and the World Food Programme.
Federated learning and zero-knowledge proof architecture, letting commercial logistics providers contribute anonymised data without breaching antitrust or data-sovereignty law.
In parallel, Layer 5’s registry is assembled at the alternate ports serving the three Phase 1 chokepoints — mills, state reserves and traders vetted with the pilot governments, with right of first refusal recorded. A matching engine is worth nothing until there is something vetted to match against.
Expand to all fourteen chokepoints. Publish open APIs so commercial TMS providers can pull macro-risk scores into their own systems, making the score a bottom-up standard.
Stranded-cargo matching goes live across the full set, with the distress formula and the fixed fee published rather than negotiated per deal.
Set out plainly, so a judge can check each row against the rest of this page.
| UNCTAD criterion | How Chokepoint Sentinel answers it |
|---|---|
| Frontier technologies | Satellite computer vision, AI digital twins, federated learning and zero-knowledge proofs — each chosen against a named alternative that fails in deployment, not selected for the word. |
| Supply chain resilience | Targets the fourteen weakest physical links in global food trade directly, moving the work from reactive crisis management to advance simulation — and closes the far side of it. When a route is lost anyway, Layer 5 places the diverted cargo with a vetted regional buyer instead of leaving it to spoil on a quay. |
| Developing country impact | The core purpose, not a secondary benefit. Designed for Low-Income Food-Deficit Countries, giving them the predictive intelligence multinationals already hold — and, when a cargo is stranded in their territory, offering it to their national reserve before any commercial trader sees it. |
| Scalability | Starts where the moral and political urgency is highest — food — and extends modularly to energy, pharmaceuticals and critical minerals on the same telemetry layer. The matching engine generalises with it: every one of those cargoes strands the same way. |
| Data readiness | Bypasses the 74% problem rather than waiting it out. A clean top-down macro-telemetry layer exists from day one and depends on no enterprise cleaning its own data. |
“When the Mississippi River floods, a US farmer gets a government subsidy. When the Black Sea closes, a family in Djibouti starves. Chokepoint Sentinel exists to give developing nations the early warning they need to act before a disruption becomes a famine.”