
FreightLine
Apify 1st Customer Award · Cal Hacks 13.0
- Team
Kuldeep Debnath
Alan Tai
Aslan Wang
Rishi Golla
- Date
- October 2026
- Event
- Cal Hacks 13.0
An autonomous back office for freight finance. Nine AI agents reconcile invoices, catch fraud, and chase payments for a freight brokerage, with every dollar traced to its source.
Overview
Every freight load generates a pile of paperwork: a rate confirmation, a bill of lading, proof of delivery, GPS pings, emails, carrier invoices, and bank deposits. These sources rarely agree. FreightLine is an agentic finance team that does this reconciliation on its own. It fills the payables clerk, fraud analyst, and collections seats at a simulated 40-person brokerage, running on an ASUS GX10.
Challenge
Brokerages reconcile documents by hand. That's how padded detention charges, duplicate invoices, double brokering, and bank-change fraud get through. The only way to handle more freight today is to hire more people.
Approach
Intake runs on the GX10. Local models classify each document and tokenize personal data before anything leaves the machine, while deterministic code reads every dollar amount, at 163 to 170 documents per minute.
Nine specialist agents each run as their own A2A service with scoped tools and permissions, and delegate to each other directly. A case waiting on a document parks, then resumes days later exactly where it stopped.
Money rules live in the database as SpacetimeDB reducers, not prompt instructions. Journals must balance to the cent, the period must be open, and the agent that requests a payment can't approve it. A SHA-256 hash-chained audit trail links every ledger cell to its source document, agent, and policy. A quality-review agent re-checks every output and writes versioned lessons that a human can roll back.
Obstacles
Models misread money, for example treating 9151.8 and 9151.80 as different amounts.
Moved all amounts to exact code, which got 132 of 132 right.
Coordinating nine concurrent agents.
Direct delegation between agents keeps work flowing even when intake restarts.
Agent memory going stale.
Elastic handles search and SpacetimeDB holds the source of truth, so recall is fast and always current.
Results
Won the Apify 1st Customer Award at Cal Hacks 13.0. Across 1,002 loads ($1.96M in invoices), the agents found $271K that shouldn't have gone out as billed. They caught 45 of 45 duplicates, 3 of 3 double-brokering cases, 15 of 15 factoring mismatches, and 4 of 4 math errors, with 0 false flags on 83 clean invoices. 0 of 1,215 planted personal values leaked, a 14,809-item backlog drained after a router restart with no loss, and the system is covered by 800+ automated tests.
Features
- Real phone calls through Cartesia, where the agent defends its detention math against a dispatcher pushing back.
- FMCSA carrier verification through Apify.
- Live dashboard with full decision traces.
- SHA-256 hash-chained audit trail from every ledger cell to its source.
- On-device intake that tokenizes personal data before it leaves the machine.
Tech stack
- Python
- FastAPI
- TypeScript
- Next.js
- React
- SpacetimeDB
- Elasticsearch / Elastic Cloud
- A2A SDK
- vLLM
- Qwen
- Nemotron
- OpenAI
- Apify
- Cartesia
- PostgreSQL
- Redis
- ASUS GX10