Rethinking Collaborative Development: A Talk at Tokyo Tech Leads Circle
Keenan Thompson and Paul Consalvi presented a live case study on turning ambiguous ideas into mission-aligned execution at Tokyo Tech Leads Circle Meetup #14.
On Monday, February 9th, I had the chance to present alongside Paul Consalvi—a professor at the University of Tokyo’s Graduate School of Frontier Sciences—at Tokyo Tech Leads Circle Meetup #14. Our talk, Rethinking Collaborative Development, was a live case study drawn from our real working relationship building Collabojin, a collaboration platform for research teams.
Where Collaboration Breaks
We opened with the problems everyone recognizes but few systems address. Great ideas die early because there’s no durable bridge from insight to execution. Credit gets detached from accountability—people claim authorship without visible contribution. And meaning gets lost in translation every time an idea moves from one person’s head to another’s hands.
These aren’t just interpersonal frictions. They’re design failures. And they’re exactly what Paul and I ran into when we started working together.
How We Actually Worked
Paul brought critical thinking and vision. He defined what “meaningful collaboration” means, challenged assumptions, and wrote ideas into concrete documents. I handled system design and implementation—asking clarifying questions constantly, converting principles into product behavior, and using AI agents heavily with human review at every step.
This wasn’t a theoretical exercise. The workflow we developed to build the product together became the product’s design pattern.
Communication Is Lossy
One of the key insights from our collaboration: verbal communication compresses meaning. Paul has a vision. He puts it into words. I form a picture from those words. That picture is never identical to the original vision.
The fix was simple but counterintuitive for a fast-moving project—write it down. I asked Paul to document his vision in detail. That document became shared context for both of us and for the AI agents assisting the build. His aspiration now lives in the codebase alongside the code.
Compound Engineering in Practice
We structured our work around the compound engineering cycle: brainstorm, plan, work, review, compound. In the brainstorm phase, we surface ambiguity before touching code. During planning, agents research the codebase, docs, and constraints. The work phase pairs AI agents with human judgment. Review happens from multiple angles—quality, security, drift from the original intent. And in the compound step, we capture decisions so the next cycle starts smarter.
The repository becomes institutional memory, not just source files.
The Product Mirror
Something we didn’t plan but found compelling: our collaboration pattern maps directly to Collabojin’s design. Where we had vague intent in one person’s head, the platform has an initiative pipeline with explicit scoping. Where we relied on clarifying dialogue before implementation, the platform builds in onboarding friction to align contributors on why before how. Where we created written context and shared memory, the platform maintains contribution logs and transparent audit trails.
We built the tool we needed to work together, and it turns out other research teams need it too.
Collabojin’s Principles
The talk laid out four principles that guide the platform:
Stewardship over status. A student can steward a project; a professor can contribute. Roles are about responsibility, not rank.
Intentional friction. The platform slows you down just enough to align on purpose before diving into execution. This isn’t bureaucracy—it’s the minimum overhead needed to prevent misalignment from compounding.
Transparent credit. Authorship is earned through visible contribution, tracked by the system rather than claimed after the fact.
Anti-rent-seeking. The system resists credit capture by design. You can’t extract value from a collaboration without contributing to it.
Why Familiar Tools Mattered
We built on a stack we already knew: Bun, React, Drizzle, PostgreSQL, Railway, Inngest. This wasn’t a default choice—it was deliberate. When you’re coding heavily with AI agents, familiar patterns let you catch agent drift faster. You notice when something’s wrong because the patterns are second nature. Time goes to problem-solving, not tool-learning. And context compounds—each session builds on AGENTS.md, vision docs, and decision logs from previous sessions.
Try Collabojin
We introduced Collabojin at the event and were happy to see several attendees sign up for the beta on the spot. If you’re working on research coordination, multi-stakeholder projects, or any collaboration where accountability and credit matter, we’d love for you to try it. Join the Collabojin beta and we’ll reach out when your access is ready.
Takeaway
Collaboration is a design problem. Better systems, better questions, better shared memory—these are the levers. The talk was an attempt to show what happens when you treat collaboration as something to be engineered rather than something that happens by default.
Thanks to the Tokyo Tech Leads Circle community for having us. If you’re interested in continuing the conversation, get in touch.
Keenan Thompson is the CEO of Arcnem AI. Paul Consalvi is a professor at the University of Tokyo, Graduate School of Frontier Sciences.