You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
My Knowledge Project started because my own podcast needed it. I co-host an improvised Dungeons and Dragons podcast. It has been going for nine seasons now, and when you improvise a story for nine years, nobody can remember what happened in season four. Not the cast, not the audience. We needed to understand what had happened historically in the show to tell the story well, and when you're improvising that's very difficult, so we needed a wiki. I went looking for a tool and the only alternative out there was volunteer labour: the fan wiki for one popular show holds 5,861 articles built from 281,525 hand edits by 36 unpaid people. So I built the tool, and then that grew into its own product.
Here's what it does. You connect a podcast feed or a YouTube channel. The system transcribes every episode with speaker attribution, then a pipeline of AI agents pulls out the people, places and running threads, reconciles them across the whole archive, and writes an encyclopedia article for each one. The part that matters: every claim in every article links to the exact second of audio it came from. A reader never has to trust the machine. They click, and they hear it. And when the machine is wrong anyway, any reader can flag it, and that correction becomes a standing rule applied to every future episode of that show. Version history makes every change reversible.
Everybody agrees that large language models make things up and that you shouldn't trust what they generate without checking it. It's maybe the most mainstream opinion in the whole AI discourse, it's correct, and almost nobody who agrees with it actually builds accordingly. The industry treats hallucination as a model-quality problem that better training is going to fix eventually. We treat it the way aviation treats engine failure: a permanent condition you engineer around. If the text can't show you its source, the system shouldn't publish it.
The goal of this grant is to make that design pattern copyable. Three things. Publish the citation-verification layer, which is the machinery that pins every generated claim to its source second and refuses to publish claims that can't be pinned, with a write-up of how it works. Publish the correction-loop design too, including the incident that shaped it: an automated repair once damaged 154 human-written links on our own platform, and we rebuilt everything afterwards so detection and mutation are never chained together again. And open the platform up from alpha to Canadian creators broadly, processing complete back catalogues for the first cohort, because that's what generates the hard cases the verification layer needs. The same machinery works past podcasts: oral histories, civic meetings, courts, any archive where who said what, when, matters and nobody can afford a research staff.
$50,000 is roughly a year of me working on this full time. We measured the cost of fully processing an episode and drove it down to about three US dollars, so the money is mostly my time, not compute. Right now I build this alongside contract writing work that pays my rent. I write somewhere between three and four scripts a week for a large YouTube operation, and the open publications are exactly the part that a bootstrapped company keeps deferring forever. A smaller grant speeds up the same path, there's no cliff. The $10,000 minimum covers publishing the verification layer and the correction-loop design without the cohort expansion.
There are two of us. Me: I started in engineering, and as I was going through it I learned I really love the fundamentals of how things work, so I did a minor in math and then a master's degree in math. I went into neuroscience research, didn't love the culture of academia, left, did an MBA, and got recruited into consulting. I worked for RBC, Canadian Tire, TransLink, basically a bunch of companies that had complicated data problems. Then I was COO of a podcast advertising platform. We had clients from day one, and my job for three years was balancing what it costs to build a thing against what pain that thing solves for our customers. We got to cash flow positive in 2025. The whole time I was doing standup at night, which is how the podcast happened: one of the comedians I worked with loved D&D and wanted to start a Dungeons and Dragons podcast. The joke of the first few seasons was that none of us knew how to play except him. That show gets somewhere around 40 to 50 thousand downloads a month now. My co-founder Brad Gill is a developer of twenty years I've worked with for six.
The platform runs nine shows in production today with thousands of episodes converted into linked articles. Bootstrapped, no outside investment, neither of us has taken a dollar from it. In a measured head-to-head on identical audio, the leading commercial competitor attributed 464 words of one episode to an unknown speaker. We attributed every word, and we found a fifth speaker their system missed entirely. Our speech-recognition provider now wants our hardest episodes as benchmark cases.
The likeliest failure is commercial, not technical. Creators might not pay enough for the company to sustain the work, and the open publications are exactly what a cash-strapped company defers forever. That's why they're sequenced first and covered by the minimum. The second risk is that we publish the pattern and nobody adopts it. The mitigation is publishing it attached to a working product instead of as a paper. And if the whole thing fails anyway, the layer and the write-ups still exist in public, which is most of what this grant is buying.
No investment. We're bootstrapped and neither founder has taken a dollar out. Vendor credits only, AWS Activate at $10,000 USD. I applied to Emergent Ventures this week, and a TELUS small-business grant and a Canada Media Fund pre-application are in flight. That's everything.