Why I'm running Parley

I started a Kaggle research project in Q2 2026 while running two startups. Its name is Parley. It's sign-language computer vision, published in monthly notebooks, with no path to becoming a product. I want to write down why.

The short version is: I have ADD, the main work needs cognitive cover, and there is a real research area that has gone quiet enough to be worth poking. The longer version is below.

The ADD context, plainly

I have ADD. I have always had ADD. The shape it takes for me as an operator is that focus on a single venture, even a venture I love, will eventually generate the kind of grinding restlessness that ends in either a bad decision or a wasted week. The bad decision is usually a scope expansion I shouldn't have made. The wasted week is usually a project I half-finish and abandon.

The pattern I've watched in myself for twenty years is that I don't get to choose whether the restlessness shows up. I only get to choose what it lands on. If I let it land on the main work, it deforms the work. If I park it on a YouTube binge or a 1 a.m. side-tab spiral, I lose the week and the work still gets deformed by tomorrow's tiredness.

What I've actually figured out is that I can route the restlessness into a defined, narrow channel that runs in parallel to the main work, doesn't compete with it, and gives the part of my brain that needs novelty something specific to chew on. The channel has to be technical, because if it isn't I'll never spend evenings on it. It has to be far enough from the main work that it feels like a different room. And it has to have explicit rails so it can't grow into a second startup.

I don't write that as a productivity hack. I write it as cognitive management. The rails are not optional. The decompression channel is what stops the main work from breaking. I learned that the hard way over years of starting projects I shouldn't have started.

Adjacent enough to compound, distant enough to rest

Parley is sign-language CV. Quantum Caddy is cornhole CV. The two share more of the stack than most people would guess from the surface.

Both depend on pose models that find keypoints on bodies in motion. Both depend on landmark models that find finer-grained points (hands, fingers, faces, board corners, bag positions). Both depend on temporal models that aggregate frame-level information into a classification across time — a sign in Parley's case, a throw outcome in QC's case. The libraries are the same. The training intuitions transfer. The failure modes rhyme.

What's different is the data, the constraints, and the domain. Sign-language CV has large, public, well-annotated datasets. Cornhole CV doesn't; we collect our own. Sign-language CV has no real-time constraint at the research level — a notebook can take an hour to score. QC has to run on edge hardware at frame-rate. Sign-language CV is a communication problem. Cornhole CV is a scoring and analytics problem. The domain gap is wide enough that working on one feels like rest from the other.

That gap is the whole reason this works as a decompression channel. If Parley were an adjacent cornhole project, my brain would treat it as more QC and the rest function would fail. If Parley were a totally unrelated topic, the skills wouldn't transfer back. Sign-language CV sits in the narrow window where it's distant enough to feel different and close enough that the muscle memory compounds.

A field that went quiet

Sign-language CV had a moment of public attention around 2023. Google's ASL Signs Kaggle competition pulled a lot of CV practitioners into the area for a few months. The competition resolved, prizes were awarded, blog posts went up, and the public research velocity dropped sharply.

What's left in the gap is a set of questions that the competition didn't answer because the competition wasn't shaped to. The Parley project brief names four of them: the relative contribution of hand-shape versus temporal modeling, the ceiling on landmark-only classification, signer dialect robustness under leave-one-signer-out, and the failure modes in the isolated-to-continuous-sign transition. Each is a notebook. Each is a real question. None of them is the kind of question a startup would prioritize, because none of them lead to a defensible product on a short timeline.

This is exactly the shape of underinvestigation I look for. The work is publishable, the questions are honest, the answers will be useful to someone, and the area is not crowded with people racing to ship. The roadmap lays out the first five notebooks in order. The first one is dataset exploration. The next four chip at the four open questions one at a time.

I'm not the first person to ask any of these questions. I am one of the few people choosing to publish careful, single-question notebooks with error bars and failure modes in 2026 instead of in 2023. That's enough of a niche to be worth a year.

What success looks like in 12 months

I've written this down so I can re-read it when I'm tempted by something else.

In twelve months I want six to eight published Kaggle notebooks, each one answering a specific question with evidence. I want the Kaggle profile to read like the work of someone who does honest applied CV research. I want enough public credibility that when QC talks to a future investor, an ACL contact, or a hire, the Parley profile is a portfolio asset that doesn't need explanation. And I want at least one technique developed in a Parley notebook that could (in a separate decision, logged separately, per the rule in Lesson 3) flow into the QC vision pipeline.

That's it. Four bullets. The success criteria in the project brief is the source of truth and it matches what I just wrote.

What success does NOT look like

Just as important: what would count as the project drifting away from its purpose.

Kaggle medals are not the goal. I am specifically not chasing leaderboards. A leaderboard chase generates the wrong notebooks, the showy ones with marginal gains and no clean explanation of why the gain is there. The whole reason I'm doing this is to learn and to publish careful work, not to compete for ranking.

A product is not the goal. If Parley becomes a product, it competes with QC for cycles, and the premise of the whole project breaks. The "won't become a product" rule is the formal version of that commitment. The brief lists no commercial licensing discussions, no real-time production system, no mobile app, and no dataset marketplace participation, in writing, before the first notebook shipped.

Ten notebooks in a month is not the goal. The cadence in the roadmap is one per month. Going faster looks productive and is actually the same trap as the leaderboard chase: more output, less rigor, and a slow drift into a different project. One per month forces depth. One per month also fits inside evenings and weekends with QC and MHG running in the foreground.

Kill criteria, in writing, before day one

The closing thought is the one that lets me start the project at all.

The kill criteria for Parley exist in writing in the roadmap under the monthly check-in checklist. On the first of every month, I ask: was last month's notebook shipped? Did any QC [URGENT] heat pull cycles this month? Are there new questions worth adding to the roadmap? If a QC fire pulls cycles, I note it in the decision log without guilt and the notebook slips. If the slip becomes a pattern, the project goes on ice. The criteria are mechanical. The pause is not a failure.

That is why I can start. The arm exists because the rails exist. Without those rails, I would have started Parley in 2023 when the Kaggle competition was live and abandoned it three months in for the same reasons I have abandoned every other side project across two decades.

This one is different because the rules are written down. If you've watched me start and abandon projects before, that's the only sentence in this post that matters.