Thesis

Turn raw port signals into a learning system.

PortMind starts with public camera frames, vessel events, schedules, and other time-stamped signals. The near-term job is not to claim a finished model. It is to build the trusted dataset and review loop that make reliable models possible.

Origin

Montreal is the first field case. The pattern is broader.

The Port of Montreal announced a CAD 6.6M data visibility initiative, targeted for 2027: freight-flow visibility, planning, coordination, automation, and operating-cost reduction. That announcement framed the first use case: systematically collect and learn from public infrastructure signals around one port, then build a method that can travel.

Source: Port of Montreal public visibility-data announcement.

What Exists Today

The capture layer is working. The learning layer is being built.

• Collection is live. Public camera views and media records have been captured for months on low-cost cloud infrastructure.
• Context is attached. Vessel arrivals, schedules, and operating records can be aligned to the same timeline as the images.
• Labeling is underway. General-purpose model labels are being checked by human review, with mistakes routed back into the queue.
• Model work is early. The repo has scoring and experiment scaffolding, but no production model is being claimed yet.

Dataset Scale

A six-month operating memory is already forming.

From the January-June 2026 inventory snapshot; collection continues.

The Capture Layer

What the system actually sees.

Sample frames from the Port of Montreal collection, across sites, seasons, and conditions.

Why It Is Hard to Copy

Not the feeds. The accumulated structure around them.

Anyone can point a scraper at a public camera. The defensible asset is what compounds: accumulated history, organized records, verified labels, a catalogue of hard cases and known mistakes, human review trails, and a repeatable improvement routine. Each cycle makes the next one cheaper and the dataset harder to replicate.

The Capability Ladder

Each rung earns the next.

* Later rungs are future directions, not current capabilities or claims.

Recognition · Status

Recognition is the first model target, not a finished product.

The work today is to make the labels and test sets strong enough that a model score means something. Small experiments can prove the training pipeline, but the product claim only starts once the model works on held-back examples from real conditions.

Research Path

The research direction is practical: learn from the unlabeled stream, then spend human review where it matters most.

• Large image streams can teach useful visual features before every image is manually labeled.
• Human review should focus on the hard cases: low light, snow, blur, tiny trucks, and confusing container scenes.
• Forecasting needs connected signals, not isolated images: camera views, vessel events, gates, weather, and time.
• Reinforcement learning comes later. First the system needs a reliable picture of the port state and a simulator that can test choices without touching real operations.

Labeling · What We Learned

General-purpose model labels are useful, but not reliable enough to blindly scale.

• Human review is load-bearing. The first pass used OpenAI GPT-5.5 to label images, but those labels need verification before they count as trusted records.
• Other model families still need testing. Anthropic and xAI models may behave differently, but they should be measured against the same human-reviewed examples.
• Hard cases are systematic: low light, snow, motion blur, small trucks.
• Ambiguity clusters: pickups and work trucks vs. cargo trucks; parked containers; container-truck vs. bare chassis.

Automation Layer

Software makes the review loop scalable.

• Monitor collection health and data freshness.
• Propose labels, but keep human review on high-impact examples.
• Flag feeds, labels, or scores that start behaving differently than expected.
• Prioritize images where model disagreement or uncertainty is highest.
• Summarize what changed in the port timeline as the dataset improves.

Why This Compounds

Time in market is the multiplier.

History cannot be backfilled. Every day of collection, labeling, review, and scoring widens the gap between this dataset and any future attempt to recreate it. The loop is simple: collect, label, verify, test, train, and requeue the mistakes. It is designed to run continuously and improve with scale.

Near-Term Focus

Finish the labeling foundation, then train against reality.

• Close the human audit loop: turn the latest review batch into a clean report and update the labels.
• Compare model labelers: measure OpenAI, Anthropic, xAI, and open-source vision models against the same reviewed examples.
• Lock evaluation: reserve real examples from night, weather, snow, blur, and distant-truck cases.
• Train only when the target is clear: start with truck and container-truck recognition, then expand.

Closing

Learn the port.
Everything else follows.

PortMind is a long-horizon bet that a structured, accumulated record of public infrastructure activity becomes the foundation for recognition, forecasting, simulation, and eventually reinforcement-learning research.

PortMind. A data capture and learning system for public port infrastructure.