Surplus local fisheries → shelf-stable canned seafood for food banks. A multi-agent AI platform that scores surplus catch, drafts procurement terms, matches canning facilities, and routes finished product to food banks — with a human approving every step.
Agent recommends. You decide. Every agent output is a draft with a reason attached — quantity, price, route, partner. A human operator approves before anything executes, and every approval is logged to an audit trail.
tidelift.tatinc.us — two ways in:
| Entry point | What it is |
|---|---|
| /guided-demo | A self-running, narrated walkthrough of the full pipeline — no clicking required |
| /intake | Hands-on, working prototype. Run the real agent pipeline yourself on a logged catch — score → procurement counter-offer → facility match → equity route — and approve or reject the result. Every decision writes to the audit trail. |
| /architecture | Reference cards: what each of the 9 pipeline steps actually does, and exactly where Cloudflare sits (edge compute vs. Workers AI vs. KV) — for anyone checking under the hood |
A screen-recorded backup walkthrough is also in the repo in case the live site is unreachable:
apps/web/static/media/guided-demo/videos/Webpage-demo-slowed.mp4.
Honesty note: on-vessel computer-vision catch counts and thermal readings shown in the demo are labeled Simulated — trained fish-scan model is roadmap. Scoring, procurement drafting, facility matching, and equity routing are the real agent logic running live, not mocked for the demo.
Per guidance from onsite industry advisors, the team prioritized the segment between the fishing vessel
and canning/processing intake — the highest-signal, least-solved gap in the pipeline. This is now a real,
interactive flow at /intake, not a mockup:
- On-vessel catch logging — computer vision at the point of harvest captures quantity, size, and species before the pickup truck is dispatched, so the purchase/segregation decision is made with real data instead of a phone call. (Simulated in the current build — see honesty note above.)
- Cold-transport dispatch — the agent pipeline scores the lot, drafts a procurement counter-offer, and routes it, all before a truck rolls.
- Dockside CV sort — species and size graded into barcoded, reusable, easy-clean bins, shellfish kept separate from fin fish from the first touch.
- Thermal-camera surface temperature profiling at intake, flagging any bin that's drifted out of range (>4°C cold-chain threshold).
- Chem/microbiological QA flags — before and after processing, not just at the loading dock.
- Full cold-chain custody and traceability — every bin's location, temperature, and handler logged from vessel to processor.
An operator can run this pipeline end-to-end today at /intake: Approve advances the lot to
PROCUREMENT_CONFIRMED and schedules an ACCFB shipment; sub-500 lb catches take an honest HOLD path
instead of being force-fit into a shipment. The procurement/canning/routing agents below remain part of the
full vision and are already built out — this segment is the current milestone. Agent recommends. You
decide — every CV read, temperature flag, and QA hold is a recommendation an operator confirms, not an
automatic action.
53 million Americans visit food banks — 1 in 6. Yet every day, millions of pounds of surplus seafood from local fisheries spoil or get discarded because the cold chain is too expensive, too fragile, and too perishable to reach those who need it most.
Frozen dies. Canned travels.
The missing piece isn't the fish — it's the complete AI agent system that handles everything end-to-end: finding and contacting fishery suppliers (who have no web presence or public contact info), drafting procurement terms, securing canning facility partnerships, managing the full repack pipeline, and routing shelf-stable protein to food banks that have no supplier relationships and no budget for a supply chain team. Agents draft every step. A human operator approves each one.
That's TideLift. Not a dashboard. Not a tool. A complete operational system — with a person in the loop at every decision point.
Most food bank tech stops at inventory management. TideLift starts before the fish is caught.
The original insight was simple: frozen seafood has perishability risk — a single power outage or logistics delay destroys the product and the relationship. The solution was to eliminate the cold chain entirely by targeting the canning step. But that created a harder problem: no one had the supplier contacts, the canning facility relationships, or the negotiation bandwidth to make it work at scale.
TideLift's agents solve all three — drafting continuously, at a cost food banks can actually afford — but never acting alone. Agent recommends. You decide.
TideLift is a multi-agent Operations Intelligence Hub built around specialized AI agents that work in concert — from on-vessel catch to food bank shelf. Every arrow below is a draft, not an action: an operator approves each handoff before it executes.
🎣 Vessel Intake 🏭 Canning Partner 🏦 Food Bank Shelf
│ │ │
▼ ▼ ▼
┌──────────┐ draft terms ┌──────────────┐ draft route ┌──────────────┐
│ INTAKE │──────────────▶│ PROCURE & │─────────────────▶│ EQUITY │
│ AGENT │ (→ approve) │ CAN AGENT │ (→ approve) │ ROUTER │
└──────────┘ └──────────────┘ └──────────────┘
│ │
└──────────────────── ANALYST AGENT ───────────────────────────┘
(NL Q&A, briefs, dashboards)
Every agent below drafts a recommendation with a reason. Nothing executes until a human clicks approve — logged to the audit trail either way.
| Agent | Role | Themes Covered |
|---|---|---|
| 🚢 Intake Agent | On-vessel CV catch metrics drive the dispatch decision; dockside sort into barcoded bins; thermal-cam QA at >4°C | Current milestone — vessel → pickup → processing |
| 🔭 Forecast Agent | Reads harvest calendars + inbound history to predict surplus windows, price dips, and supply gaps 2+ weeks early | Theme 1 — See It Coming |
| 🤝 Procurement Agent | Finds fishery supplier contacts, drafts volume-discount terms, books canning facility slots, prepares POs for approval — no Rolodex required | Theme 3 — Production Is Manufacturing |
| 🏭 Canning Ops Agent | Treats repack as a factory line — forecasts kit needs, stages raw material, predicts volunteer turnout, sequences the floor | Theme 3 — Production Is Manufacturing |
| 🚚 Equity Router | Drafts routes on access windows + dietary need, not just miles; re-proposes routes in minutes when a truck goes down | Theme 4 — Equity With a Truck Attached |
| 📊 Analyst Agent | Answers plain-language questions from anyone on the team — no SQL, no tickets; briefs every lead at shift start | Theme 6 — Every Team Member an Analyst |
Full interactive version at tidelift.tatinc.us/architecture
TideLift covers every step a human team would otherwise need to coordinate manually:
- Vessel intake — Intake Agent scores the logged catch and flags cold-chain risk before the truck rolls
- Supplier outreach — Procurement Agent finds contacts for local fisheries (no public directory needed)
- Discount drafting — Agent proposes pricing and volume terms; an operator approves before anything is committed
- Canning facility booking — Procurement Agent identifies co-packing partners and drafts booking requests
- Canning ops management — Canning Ops Agent manages the floor: staging, sequencing, volunteer scheduling
- Finished goods routing — Equity Router proposes delivery by access window, dietary need, and urgency; operator confirms
- Continuous reporting — Analyst Agent keeps every stakeholder informed in plain language
This is the pipeline that didn't exist. Food banks had no way to work with local fisheries at scale because the middle steps — contact finding, negotiation, canning logistics — required a full supply chain team. TideLift is that team, with every recommendation surfaced for a human to approve, never a black box acting on its own.
TideLift directly addresses Alameda County Food Bank's 35-item Operations Intelligence Hub request across all 7 challenge themes:
- Produce/fish surplus forecast off harvest calendars and inbound history
- Vendor scoring on delivered vs. committed performance
- Price window flagging for purchased food before prices move
- Canning line treated as a repack factory — box/kit forecasting weeks out
- Volunteer turnout prediction and no-show risk flagging
- Raw material staging generated the night before, by station, in order
- Routing on agency access windows and product urgency, not just miles
- EV charging schedules matched to tomorrow's routes and daytime power cap
- Route rebuild in minutes when a truck goes down or agency cancels
- Dietary preference matching: inbound seafood matched to neighborhood nutrition needs
- Pull-based replenishment replacing pre-allocation guesswork
- Natural language Q&A for drivers, receivers, and warehouse floor staff
- Shift-start briefs: what's inbound, what's at risk, what changed overnight
- Voice note → structured proposal pipeline
- Cross-food-bank surplus matching before product ages out
- Anonymized benchmark data shared across the network
"Frozen has perishability — which is exactly the vulnerability we needed to eliminate."
| Frozen | Canned (TideLift) | |
|---|---|---|
| Cold Chain Required | Yes — costly, fragile | ❌ None |
| Shelf Life | ~3–6 months | 2–5 years |
| Distribution Reach | Urban hubs only | Rural, remote, any pantry |
| Waste Risk | High (power outages, transport) | Near zero |
| AI Draftability | Low — spot market | ✅ Contractable, forecastable |
| Supplier Discovery | Manual outreach required | ✅ Agent-drafted, human-approved |
Canning transforms a volatile, time-sensitive commodity into a predictable, distributable asset — exactly what food bank supply chains need. And because agents draft supplier discovery and procurement terms for a human to approve, food banks don't need existing fishery relationships to start.
Frontend: SvelteKit 4 (real-time dashboard, vessel intake, logistics kanban, audit trail)
Deploy: Cloudflare Pages (adapter-cloudflare), KV-backed store
Agents: TypeScript modules — intake, score, draft-procure, match facility, route, analyze
Backend: TypeScript / Node.js, pnpm workspace, CI-tested (vitest)
Data: Harvest calendars, USDC fishery data, inbound history APIs
AI: LLM-powered drafting, supplier discovery, and NL Q&A — human approves every output
Every agent is plain deterministic TypeScript — 7 of the 9 pipeline steps never call a model at all. Workers AI is used only to phrase two rationale sentences (opportunity score, procurement outreach copy), with an 8-second timeout and a deterministic fallback so a model hiccup can never break a live demo. Cloudflare KV is the one durable store behind the operational app — lots, approvals, shipments, and the audit log.
/
├── apps/
│ ├── web/ # SvelteKit dashboard — vessel intake, kanban, partners, audit trail, approvals API
│ └── agents/ # Agent modules: intake, scorer, procure, canning, route, analyst + vitest tests
├── packages/
│ └── shared/ # @tidelift/shared — shared types, schemas, mock data
└── docs/
├── PITCH.md # Deck and one-pager
├── ARCHITECTURE.md # Agent flow diagrams
├── DEMO_SCRIPT.md # Demo script and walkthrough
└── ACCFB_NUMBERS.md # Quantified ACCFB problem/solution figures
# Install dependencies (pnpm workspace — see pnpm-workspace.yaml)
pnpm install
# Run the dashboard
pnpm --filter web dev
# Run agent tests
pnpm --filter agents test- 🐟 US fisheries discard millions of tons of bycatch and surplus annually
- 🥫 Canned seafood is one of the most requested but least donated food bank items
- 📋 No supplier directory exists — food banks can't find fisheries, fisheries can't find food banks
- 🤖 No AI system currently exists to draft the fishery → procurement → canning → food bank pipeline end-to-end for a human to approve
- 💰 $2,500 First Prize — AISCO Hackathon 2026 (Judging: July 17, Capgemini)
| Criterion | TideLift's Answer |
|---|---|
| Solves a real problem | Directly from ACFB's 35-item build list + field research |
| AI agent-driven | Specialized agents with defined roles and handoffs, running live at /intake |
| Quantifiable impact | Surplus fish → shelf-stable protein, cold chain eliminated, 2–5yr shelf life |
| Human in the loop | Agent recommends. You decide. Every draft carries a reason; an operator approves before execution; every approval is logged to the audit trail |
| Built for Alameda County | All 7 themes addressed, ACFB operations specifically modeled |
| Complete system | Vessel intake + procurement drafting + canning logistics + routing — not just a dashboard |
Built at the AI Supply Chain Observatory Hackathon 2026 — Food Banks + AI Hosted by the AI Objectives Institute (AOI) July 15–17, 2026 | Pebblebed → Capgemini
TideLift — because the tide lifts all boats, and every neighbor deserves protein on the shelf.