Weekly Update · July 14-21, 2026

This week, PortMind turned review work into a usable product.

An invited reviewer can now inspect a port image, record a decision, and save the result. The next step is to finish the first checked set and take it into customer conversations.

What Moved

Three concrete steps moved the project forward.

01 · SHIPReviewer appInvites, instructions, image inspection, decisions, and saved notes now work in one flow.
02 · PACKAGEFirst reference setThe immediate deliverable is a small, checked set of port images that we can show and test.
03 · ALIGNTwo callsCounsel today; Simon Savard at the Port of Montreal tomorrow.

The focus is execution: finish the first set, learn from it, and use the result to guide the next customer conversation.

Roadmap

Build the port-data standard, then turn it into products.

01 · CollectPort datasetAdd time-stamped images and events from each port we serve.
02 · LabelPort assetsMark tankers, trucks, vessels, and hard cases with reviewer checks.
03 · BenchmarkPort tasksCompare systems on detection, classification, and bounding boxes.
04 · ReleasePublic + privateShare a public benchmark; sell private evaluations and labeling.
05 · ApplyPort toolsUse the history to build monitoring, alerts, and forecasts.

The open release is a defined benchmark and supporting tooling; private customer work stays private. The first benchmark starts with five systems and expands as review capacity grows.

Evaluation

Measure every signal, with a human in the loop.

Established benchmark72 hard-case rows
historical comparison
LabelerTruckContainer truck
Human reference
72 rows		ReferenceReference
Codex
72 rows		59.7%50.7%
Grok 4.5
72 rows		77.6%67.2%

Agreement against the human reference. Useful as a first-pass signal; human review remains required.

This week’s open-model pass24 calibration rows
agreement vs Human 1
ModelTruckContainer truck
Llama 3.2 Vision95.7%82.6%
LLaVA 1.5100%82.6%
Moondream 3.1100%78.3%
Llama 4 Scout26.1%91.3%
Mistral Small 3.1100%87.0%
Second human labeler9 / 16 known-answer controls

56.3% scene agreement; consistent, but failed qualification. Retrain before promotion.

Open-model scores use the same 24-row packet as Human 2; binary scores use 23 completed reference rows. Report-only calibration results, not production accuracy or benchmark truth.

Human Review

The reviewer app now turns an image into a saved, checkable decision.

PortMind Reviewer profile setup screen
StartGive each reviewer a clear assignment and set of instructions.
PortMind Reviewer scene decision screen
DecideShow the image, the question, confidence, and notes before saving.
PortMind Reviewer inspection grid view
InspectUse the grid to make difficult images easier to check consistently.

Screenshots from the PortMind reviewer workspace. This is the app itself, not the onboarding video shown in the previous update. Next: complete the first checked set.

The Review Loop

Each image follows the same path from invite to usable record.

  1. InviteSend a reviewer a scoped link to the assigned work.
  2. Set the rulesExplain what to look for and what a good decision includes.
  3. InspectUse the original image and grid view to check difficult scenes.
  4. RecordSave the decision, confidence, image quality, and a short note.
  5. PackageMove checked records into the reference set and customer demo.

Next Conversations

Use the next two calls to remove the next blockers.

Today · Counsel

Choose the structure

Leave with a clear project structure, ownership, and the next workstream to execute.

Tomorrow · Port of Montreal

Validate the first use case

Ask Simon Savard which port workflow is worth testing first and who should evaluate it.

Role reference: Port of Montreal Innovation Hub. Portrait: Startupfest public speaker profile.

Next Seven Days

Finish the first set, then use it to make the next sale.

  • Complete the reviewer flow. Open the first set and get every assigned image reviewed.
  • Summarize the results. Record where reviewers agree, disagree, or need clearer rules.
  • Prepare the customer demo. Show the checked set and one practical port use case.
  • Close the loop on both calls. Leave with owners, dates, and one next test.