Weekly Update · July 21-31, 2026

This week, PortMind turned into a real research track.

We tightened what the project is, built a better way to test models, prepared the Oloodi incubation agreement, and opened the next Port of Montreal conversation.

What Moved

Four major pieces landed.

01 · DEFINEClear rulesWrote down what counts as a truck, container truck, unclear image, and valid answer.

02 · BUILDModel testingBuilt a cleaner way to test models and record how they reached each answer.

03 · EXPLAINSimple storyUpdated PortMind’s public-facing message so the project is easier to understand.

04 · ALIGNOloodi pathDrafted a simple agreement for PortMind to be incubated inside Oloodi.

The short version: PortMind is now easier to test, easier to explain, and closer to having an operating home.

Research Foundation

Before scoring models, we first made the task clear.

01 · QuestionAskWhat are we asking the person or model to decide?

02 · RulesDefineWhat counts as yes, no, or unsure?

03 · HumanReviewHave people check the hard images first.

04 · ModelTestRun each model on the same images and rules.

05 · ResultReportShow what worked, what failed, and what still needs review.

Model Evaluation

We learned that not every model can be tested the same way.

Image-only laneLow-cost
open models

Test type	Models used	What it means
Image only	Llama 3.2 Vision, LLaVA 1.5, Moondream 3.1	The model sees the image and gives an answer.
More open models	Llama 4 Scout, Mistral Small 3.1	Useful for cheap baseline comparisons.
Prepared evidence	Same open-model lane	We can give the model crops or zoomed views, but the model is not choosing tools by itself.

This is the practical lane for Cloudflare and other lower-cost vision models.

Tool-using laneHigher cost
agentic models

Test type	Models used	What it means
Model chooses tools	Kimi K3 through OpenCode	The model can ask to zoom, crop, or inspect before answering.
Agent system	Grok 4.5 through Grok Build	A larger agent harness helps run the work, so it should be scored separately.
Paid model packet	None this week	Use top paid models only after the test set and budget are ready.

Smoke resultKimi K3 used tools

The test proved we can record tool calls. It also showed the answers still need better rules and review.

The main learning: answering the image and choosing useful tools are two different skills. PortMind now tracks them separately.

Brand

PortMind is still the right name to build around.

What we foundSearch check
July 31

Result	What it is	Why it matters
portmind.com	Trade-compliance AI company, founded in Geneva and operating from Dubai.	Name overlap, but different focus.
Dubai CommerCity news	Portmind partnership around customs and trade documentation.	Shows the name has existing AI/logistics use.
portmind.ai	Domain acquired for this project.	Clean AI positioning for the research and benchmark story.

Sources checked: portmind.com, Portmind company page, and Dubai CommerCity announcement.

Oloodi

The proposed agreement is simple: incubate PortMind inside Oloodi.

PortMind originWielfried created the idea, process, and work so far.
Oloodi roleOloodi becomes the Montreal operating home while the project is early.
Initial supportOloodi provides affiliation, business guidance, introductions, and a proposed CAD 750 August research budget.
Ownership principleWielfried keeps founder leadership; Oloodi receives real upside if it incubates and funds the project.
Next documentA short ownership term sheet comes before large funding, incorporation, a paid pilot, or outside investment.

Outreach

Simon replied. The next contact is Chad McLarnon.

Port of Montreal

Simon Savard

Sent the PortMind note and received a response. The thread now moves to a strategy-side contact.

Next email

Chad McLarnon

Director, Strategy at the Montreal Port Authority. Public Port of Montreal notes describe his work around strategy, business transformation, change management, and strategic-plan execution.

Reference: Port of Montreal, September 2023 team update.

Current State

PortMind now has three tracks that are easier to explain.

RESEARCHTrusted test setBuild a small set of hard images with clear human answers.

PRODUCTReviewer workflowMake it easy for people to review images and save clean decisions.

PARTNEROloodi incubationThe project now has a proposed operating home, budget ask, and ownership conversation.

The next move is straightforward: build the trusted image set, keep cheap model tests running, and save expensive model tests for a funded evaluation packet.

Next Seven Days

Turn the new structure into action.

Send Chad the right note. Keep it short, research-first, and tied to strategy fit.
Review the Oloodi PDF. Get Boris’s reaction to incubation, budget, and ownership terms.
Continue the reviewer pipeline. Move toward the first clean human-reference packet.
Keep model runs scoped. Use open models for cheap baseline learning; hold top paid models for budgeted evaluations.