You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Large language models make things up, and text they generate should not be trusted without verification. Nearly everyone agrees with this, and almost nobody who agrees builds accordingly. My Knowledge Project is a working system built on the opposite default: if the text cannot show you its source, the system should not publish it.
Connect a podcast feed or YouTube channel. The system transcribes every episode with speaker attribution, then a pipeline of AI agents extracts the people, places and running threads, reconciles them across the whole archive, and writes an encyclopedia article for each. Every claim in every article links to the exact second of audio it came from. A reader never trusts the machine: they click, and they hear it. When the machine is wrong anyway, any reader can flag it, and the correction becomes a standing rule applied to every future episode of that show. Version history makes every change reversible.
The industry treats hallucination as a model-quality problem that better training will fix. We treat it as a permanent condition to be engineered around, the way aviation treats engine failure. That design stance, provenance as architecture rather than disclaimer, is cheap to copy once somebody shows the working pattern, and showing the pattern is the goal.
Three deliverables: publish the citation-verification layer (the machinery that pins every generated claim to its source second and refuses to publish claims that cannot be pinned) with a design write-up; publish the correction-loop design, including the incident that shaped it (an automated repair once damaged 154 human-written links, and we rebuilt the system so detection and mutation are never chained again); and open the platform from alpha to Canadian creators broadly, processing complete back catalogues for the first cohort, which generates the adversarial cases the verification layer needs. The machinery generalizes past podcasts to oral histories, civic meetings, courts, any archive where who said what, when, matters.
$50,000 funds roughly a year of focused full-time work on the three deliverables above. We measured and drove the cost of fully processing an episode to about three US dollars, so most of the money is my time, not compute. I currently build this alongside contract writing work that pays my rent, and the open publications are exactly the part a bootstrapped company defers forever. A smaller grant accelerates the same path; there is no cliff. The $10,000 minimum covers publishing the verification layer and the correction-loop design without the cohort expansion.
Two of us. Me: engineering degree, master's in mathematics, analytics for RBC, Canadian Tire and TransLink, then COO of a podcast advertising platform I helped take from launch to cash flow positive. I also co-host a nine-year improvised D&D podcast with about 45,000 downloads a month, which is where the problem came from: improvise a story for nine years and nobody, cast or audience, can remember season four. The only existing alternative is volunteer labour; the fan wiki for one popular show holds 5,861 articles built from 281,525 hand edits by 36 unpaid people. My co-founder Brad Gill is a developer of twenty years I have worked with for six.
The platform runs nine shows in production today, with thousands of episodes converted into linked articles. Bootstrapped, no outside investment, no salaries taken. 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 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 may not pay enough for the company to sustain the work, and the open publications are exactly what a cash-strapped bootstrapped company defers forever. That is why the publications are sequenced first and covered by the minimum. A second risk is that the verification pattern gets published and adopted nowhere; the mitigation is publishing it attached to a working product rather than as a paper. If the project fails entirely, the layer and the write-ups still exist in public, which is most of what this grant is buying.
No investment raised; the company is bootstrapped and neither founder has taken a dollar from it. Non-dilutive vendor credits only (AWS Activate, $10,000 USD). An application to Emergent Ventures went in this week, and a TELUS small-business grant and a Canada Media Fund pre-application are in flight. Nothing else.