Activity

Contributed the observed conversation handle to MCP group results, preserving message permissions and account boundaries. Merged upstream in PR #179.

Contributed guidance that distinguishes connection state, chat availability and history coverage. Merged upstream in PR #180.

Released evidence-based resume comparison with validated excerpts, bounded uploads and server-enforced usage budgets.

Integrated Coworking Scout into the public map, with coworkings, cafés and libraries from an approved public catalogue.

Put the open resume matcher on transactional Postgres rate limits with authenticated daily cleanup, added DOCX decompression preflight, bounded public model attempts, and kept advisory resume evidence separate from operational applicant scoring.

Turned raw Places results into a working queue with ready-to-call prioritisation, filters, sorting, search and responsive pagination, and moved CRM timestamps onto the database clock. This is a pilot, not a self-service product.

The mobile sidebar set touch-action to none but only listened for clicks, so every drag gesture was dead. Rebuilt it as a real bottom sheet with peek, mid and expanded anchor points, where flick momentum projects the landing anchor.

Contributed voice-command improvements to a shared Chrome extension: percentage scrolling, search highlighting and a scroll indicator.

A direction for Voidscape: strengthen the installed CLI and recovery workflow first, then explore richer review interfaces. This is a roadmap, not a released feature.

From a talk on AI governance at Trajectory Labs in Toronto. Spanish and Portuguese make up barely three percent of the datasets behind top frontier models, and when a model ignores your culture the software misses local legal rules, regional dialects and everyday context. That is why regional models like LatamGPT matter, and why education work like BrainTrainr makes a direct impact across Latin America and other underrepresented groups. Governance is not abstract theory: it decides who receives compute access, who sets safety rules, and who holds a seat at the table.

Across Startupfest and conversations with founders building in the agent space, the same challenge kept showing up from different angles. Once an agent can take action, the question is no longer whether it generates good code, but whether anyone designed clear boundaries around what it is allowed to do. Kastra.ai reviews risky actions before they execute. Jetty replaces open-ended prompts with runbooks that define the task and what done looks like. Upivia treats agent access like a controlled wallet, a policy-checked gateway with logs and budget limits. Same shift: permissions, workflow constraints and memory get designed on purpose, rather than added after something breaks.

The best AI tools make setup easy so you can start right away. MigmaAI can begin from a website, a prompt, a Figma frame or an existing email and build something editable that matches the brand. What is interesting is not the speed: email is tricky, and a design that looks fine in an editor can break in Outlook, Gmail, dark mode or on mobile, so the product still has to let people review, edit and test what lands in the inbox. I do not think good AI products should remove interfaces entirely. People need control, context and a way to fix things. I am more interested in AI that takes away setup steps than in AI that adds another chatbot to a workflow.

At an event a few weeks ago, two people sat down and led with their company name, then their school. That was the whole pitch, no questions back, no curiosity about anyone else at the table, and they left before anything real got said. The ones who stayed asked what you were building, what you were stuck on, what you cared about outside work. Same room, same free drinks, completely different night. So my test is not how many connections I can collect. It is how many people I would still grab a coffee with if neither of us could do anything for the other.

Working from coworking spaces, I shipped more, stayed sharper and felt less isolated. The social layer of building is not a distraction, it is fuel. I still like the convenience of working from my room, but sometimes the time you optimize comes at a cost: it gets dull, there is no stimulus, and doomscrolling gets easier. I was not replenishing energy, I was compressing and getting drained. This is the itch behind Coworking Scout on PeruGrid.