Document Intelligence
DODA reader
- In
- A DODA as PDF or photo
- Out
- Seventeen structured fields
Runs in DODA SmartIn production
AI Lab
The research laboratoryof Freight Technologies.
We are the AI Lab, where many of Freight Technologies’ innovations happen: we apply the scientific method to the design and development of logistics solutions, and we research together with universities. What we publish is recorded here.
Capabilities
AI capabilities,composed foryour business.
Modular intelligence.
We develop different AI-based algorithmic tools and assemble them together to solve real problems. From data to science, and from science to decisions.
An animation: capability tiles drop onto a board one at a time and lock around a tile in the middle, which is the finished product; lines then connect it to the systems a business already has.
All together
Logistics, complete.
Pricing, capacity, customs, tracking, proof of delivery, payment: every system the operation runs on is composed from the same capabilities. Put all of them together and what you get is not one more product — it is the whole of logistics.
51 pieces · 7 capabilities
Price a lane and cover it
Goes inA lane and a trailer type
- 01Lane ratePricing EngineAn estimated rate and the road distance
- 02Carrier recommenderGraph AnalyticsA ranked shortlist, with the reasons
- 03Carrier searchAI ModelsCarriers that run it, by how well they fit
- 04Capacity checkAI ModelsWhether it can run it, and whether it can cross
- 05Rate negotiatorAI AgentsAccept, counter or decline
Comes outA carrier, at an agreed rate
Clear the border
Goes inA photo of a DODA
- 01DODA readerDocument IntelligenceSeventeen structured fields
- 02QR finderComputer VisionThe SAT’s own link for that document
- 03SAT status trackingData LayerIts state, normalised, and its pedimentos
- 04Change noticeData LayerThe full record, sent to the system that asked
Comes outThe unit’s customs state, as it changes
Watch the trip
Goes inA truck’s GPS signal
- 01Unified positionData LayerOne normalised position
- 02Road ETAAI ModelsThe expected arrival, by road
- 03Rule alertsData LayerTwelve kinds of alert: stop, delay, detour and more
- 04Executable protocolAI AgentsAutomatic steps, a human signature, an audit trail
Comes outAn exception, already being handled
Prove the delivery and release the payment
Goes inA scanned proof of delivery
- 01Page straightenerComputer VisionThe page, upright
- 02Stamp and signature detectorComputer VisionEach stamp and signature, located and read
- 03Rejection detectorComputer VisionAny sign of rejection, return or damage, printed or handwritten
- 04Delivery verdictAI ModelsApproved, review or rejected, with a confidence
- 05Three-way freight auditData LayerA payment that can be released, or the reason it cannot
Comes outA payment that can be released
Built as capabilities. Composed as solutions.
The AI Lab builds modular intelligence for real-world operations. Combine specialized capabilities into a platform tailored to your needs, and connect each one to your existing systems through its own API. From pricing and AI agents to computer vision and document intelligence, our capabilities work independently or together as part of a larger solution.
Each capability is designed to solve a specific problem, connect with the tools and systems you already use, and work alongside other components in the platform.
- 01Modular by designSelect and combine capabilities around your business needs.
- 02API-firstIntegrate each capability with your existing systems and workflows.
- 03ComposableStart with one component and extend the solution over time.
- 04Purpose-builtApply the right combination of models, data, and tools to each use case.
The catalogue
What is on the shelf today.
Sorted by capability. Pick one up to see what it takes, what it hands on, where it runs and how far along it is.
Pricing Engine 6
What a lane costs, how that has moved, and which lanes resemble it.
AI Agents 13
Take a question or a task and carry it through, in chat or by email.
AI Models 8
Trained on the company’s own operating record to match, score and predict.
Data Layer 10
Turn many sources into one thing a system can act on.
Computer Vision 5
Find what is on a page: stamps, signatures, codes, which way is up.
Graph Analytics 2
Read the network of carriers and lanes to recommend and to classify.
Document Intelligence 7
Read a document the way a person would, and hand back fields.
Computer Vision
QR finder
- In
- A customs document
- Out
- The SAT’s own link for that document
Runs in DODA SmartIn production
Document Intelligence
Invoice and proforma reader
- In
- An invoice as PDF or image
- Out
- Its line items, structured
Runs in DODA SmartIn production
Document Intelligence
Goods check
- In
- A document and the goods declared
- Out
- Whether they agree, and why not
Runs in DODA SmartIn production
Computer Vision
Stamp and signature detector
- In
- A page of a proof of delivery
- Out
- Each stamp and signature, located and read
Runs in POD ValidatorIn development
Computer Vision
Rejection detector
- In
- A page of a proof of delivery
- Out
- Any sign of rejection, return or damage, printed or handwritten
Runs in POD ValidatorIn development
Computer Vision
Placeholder detector
- In
- A page
- Out
- A real proof of delivery, or a blank stand-in
Runs in POD ValidatorIn development
Computer Vision
Page straightener
- In
- A scanned page, any way up
- Out
- The page, upright
Runs in POD ValidatorIn development
Document Intelligence
Key-field reader
- In
- An image of a delivery document
- Out
- Reference, carrier, pickup and invoice numbers
Runs in POD ValidatorIn development
Document Intelligence
Availability reader
- In
- A carrier’s message, text or photo
- Out
- Where, which unit, when and how many
Runs in SpotsIn internal use
Document Intelligence
Carrier file intake
- In
- Documents a carrier sends by email
- Out
- Each one filed by type, without duplicates
Runs in SpotsIn development
Pricing Engine
Lane rate
- In
- Origin, destination, equipment and month
- Out
- An estimated rate and the road distance
Runs in ZayrenIn production
Pricing Engine
Bulk pricing
- In
- A spreadsheet of lanes
- Out
- A rate and up to five carriers for each
Runs in ZayrenIn production
Pricing Engine
Lane statistics
- In
- A lane and its equipment
- Out
- City-to-city and state-to-state history
Runs in ZayrenIn production
Pricing Engine
Rate trend
- In
- A lane and a period
- Out
- How its rate has moved
Runs in ZayrenIn production
Pricing Engine
Similar lanes
- In
- One lane
- Out
- The lanes that behave like it, priced
Runs in ZayrenIn production
Pricing Engine
Border-crossing analysis
- In
- A cross-border lane
- Out
- How each crossing behaves
Runs in ZayrenIn production
AI Models
Carrier search
- In
- A lane, equipment, weight and volume
- Out
- Carriers that run it, by how well they fit
Runs in ZayrenIn production
AI Models
Capacity check
- In
- A carrier and a lane
- Out
- Whether it can run it, and whether it can cross
Runs in ZayrenIn production
AI Models
Compliance and risk score
- In
- A carrier
- Out
- A compliance score and a risk scorecard
Runs in ZayrenIn production
AI Models
Carrier verification
- In
- Tax and authority numbers, by email or chat
- Out
- Verified, or exactly what is missing
Runs in SpotsIn internal use
Graph Analytics
Carrier recommender
- In
- A lane, equipment and a budget per mile
- Out
- A ranked shortlist, with the reasons
Runs in Carrier recommenderIn development
Graph Analytics
Carrier archetypes
- In
- A carrier’s history
- Out
- What kind of carrier it is
Runs in Carrier recommenderIn development
AI Models
Road ETA
- In
- A position and a destination
- Out
- The expected arrival, by road
Runs in Fr8RadarIn development
AI Models
Delivery verdict
- In
- The evidence from every page
- Out
- Approved, review or rejected, with a confidence
Runs in POD ValidatorIn development
AI Models
Shipment match
- In
- Fields read from the page and the shipment record
- Out
- Which fields agree, and on what evidence
Runs in POD ValidatorIn development
AI Models
Tariff code suggestion
- In
- A description of the goods
- Out
- A suggested code and its grounds
Runs in DODA SmartIn development
AI Agents
Rate negotiator
- In
- A lane and a carrier’s offer
- Out
- Accept, counter or decline
Runs in ZayrenIn production
AI Agents
Carrier analyst
- In
- A question, in plain language
- Out
- A comparison, a ranking, the contacts
Runs in ZayrenIn production
AI Agents
Spend analyst
- In
- A question about a shipper’s history
- Out
- Where the money went, and where it can be saved
Runs in ZayrenIn production
AI Agents
Customs document agent
- In
- A DODA and a question
- Out
- An analysis, its anomalies, what to do next
Runs in DODA SmartIn development
AI Agents
Customs file copilot
- In
- A customs file and a question
- Out
- An answer grounded in that file
Runs in DODA SmartIn development
AI Agents
Operations copilot
- In
- A question about the operation
- Out
- An answer with the shipment’s context
Runs in Fr8RadarIn development
AI Agents
Alert resolver
- In
- An alert
- Out
- A diagnosis and the next action
Runs in Fr8RadarIn development
AI Agents
Executable protocol
- In
- A trigger
- Out
- Automatic steps, a human signature, an audit trail
Runs in Fr8RadarIn development
AI Agents
Verdict writer
- In
- A validation result
- Out
- Two sentences a person can act on
Runs in POD ValidatorIn development
Data Layer
Chat group listener
- In
- Messages from carrier groups
- Out
- Availability and activity, as they are posted
Runs in SpotsIn internal use
AI Agents
Campaigns and cadences
- In
- A segment of carriers and a lane
- Out
- Scheduled messages, and who answered
Runs in SpotsIn internal use
AI Agents
Negotiation by chat
- In
- A request for a quote
- Out
- The conversation, its state and the closing contact
Runs in SpotsIn internal use
AI Agents
Lane outreach
- In
- A lane
- Out
- A message to every carrier that runs it
Runs in ZayrenIn production
Document Intelligence
Document inbox
- In
- An email with attachments
- Out
- Each document read and filed to its operation
Runs in DODA SmartIn development
AI Agents
One operational thread
- In
- Chat, email and voice about a shipment
- Out
- One conversation per shipment
Runs in Fr8RadarIn development
Data Layer
Change notice
- In
- A change of state
- Out
- The full record, sent to the system that asked
Runs in DODA SmartIn production
Data Layer
Unified position
- In
- Signal from 92 GPS and telematics sources
- Out
- One normalised position
Runs in Fr8RadarIn production
Data Layer
Rule alerts
- In
- Live telemetry
- Out
- Twelve kinds of alert: stop, delay, detour and more
Runs in Fr8RadarIn development
Data Layer
Geofences and route check
- In
- A position and the planned corridor
- Out
- Inside, outside or off route
Runs in Fr8RadarIn development
Data Layer
SAT status tracking
- In
- A DODA with its QR
- Out
- Its state, normalised, and its pedimentos
Runs in DODA SmartIn production
Data Layer
CFDI check
- In
- A CFDI
- Out
- Valid or not, against the SAT
Runs in DODA SmartIn development
Data Layer
Availability map
- In
- The availability captured
- Out
- Supply by origin and day, demand by lane
Runs in SpotsIn internal use
Data Layer
Location search
- In
- Free text
- Out
- A city, state and country, normalised
Runs in ZayrenIn production
Data Layer
Three-way freight audit
- In
- The rate, the carrier’s invoice and the proof of delivery
- Out
- A payment that can be released, or the reason it cannot
Runs in Fr8RadarIn development
These pieces run inside our products. They are not offered as public APIs yet; that platform is being built.Talk to our team
Negotiation
Researched, open, and playable.
The lab studied how AI agents negotiate when they compete for scarce loads. It published the code so anyone can test their own agent, and turned the problem into a game.
The lab’s game
We Have a Deal
Uruapan, 2150. Enrique trusts his instinct; Humberto bets on artificial intelligence. Nine loads of avocado each, twelve months, and carriers to negotiate with, round by round. The better negotiator wins.
- A 3D world in the browser
- Up to three rounds with each carrier
- Rates and rivals are simulated; your game is saved only in your browser
Play The game’s code on GitHub
Plays in any browser with WebGL2, with touch controls on a phone; best on a larger screen. About 14 MB.
The study · Preprint
A tournament of AI negotiators, under scarcity.
Nine negotiation engines — rules, search, optimization, imitation and reinforcement learning — and nine language models, all negotiating at once on both sides of one freight market. A stable matching has the last word.
62% → 16%
Match rate of the reinforcement-learning agent trained on isolated negotiations, once shippers were scarce and it had to compete for them.
> 91%
The agent trained against rivals kept its match rate above this in all three market conditions.
32 of 36
Comparisons nine language models, from 1 to 72 billion parameters, lost to the two strongest engines.
From the preprint “Before AI Agents Negotiate for Us”, AI Lab, Freight Technologies, November 2026.
The toolkit · Open source
Bring your own negotiator.
We Have a Deal is an open arena: write a shipper or carrier agent, plug it in, and benchmark it against the nine reference engines from the study. Everything runs on your own machine.
- Nine reference engines
- The paper’s data, figures and scripts
- Apache-2.0
The study · Preprint
Voice agents that negotiate by phone, timed from the caller’s ear.
Three ways to build the same agent — two cascades orchestrated by Pipecat and by LiveKit, and GPT-Live, a model that listens and speaks — negotiating freight in Spanish, 77 calls each, measured outside the frameworks.
13% vs 3%
Error on spoken amounts — the words that cost money in a negotiation — in the two cascades and in GPT-Live.
3 · 2 · 0 of 15
Calls in which an “ok” said over the agent ended in a registered agreement: Pipecat, LiveKit, GPT-Live.
From the preprint “Latency, Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation”, AI Lab, Freight Technologies, November 2026. Counts in the last panel are descriptive, not tested.
The record
What the lab publishes.
01
Papers
Research papers by the lab and the people it works with.
- Latency, Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation: An Externally Measured Comparison of Pipecat, LiveKit and GPT-LiveJ. Popoca Moran, E. de Ávila-Armenta, R. A. Aquino-Ibarra, J. O. Nicolás-Díaz, U. León-Domínguez · Preprint · 12 November 2026 · PDFThree ways to build a voice agent that negotiates freight by phone in Spanish, measured from the listener’s side over 77 calls each. They differed less in speed than in how they heard spoken amounts and handled an “ok” said over them. The preprint is being prepared for arXiv.
- Before AI Agents Negotiate for Us: A Competitive Tournament of Rule-Based, Search, Optimization, Learned, and Language-Model Negotiators Under ScarcityE. de Ávila-Armenta, R. A. Aquino-Ibarra, J. O. Nicolás-Díaz, U. León-Domínguez · Preprint · 11 November 2026 · PDFNine negotiation architectures, from hand-written rules to reinforcement learning, and nine large language models, negotiating freight prices against each other in one market with stable matching, under balanced supply and under scarcity on either side. The preprint is being prepared for arXiv.
02
Publications
Articles, reports and other published work.
Under review.
03
Toolkit
Tools and resources the lab makes available.