Swisspreneur runs a long-form interview podcast, a Slack community of 700+ founders, and an investment syndicate that has channelled CHF 3M+ into startups.
The show is the centre of gravity. 570+ episodes of candid conversations with founders about fundraising, hiring, scaling, and failure.
Swisspreneur had a deep archive and a real question: was there something genuinely useful to do with AI here, or was it just noise? The constraints were set before the first line of code. Weeks, not quarters, and a community-org budget.
570+ episodes hold years of operating context, and the format keeps all of it invisible. None of it can be queried until it's transcribed and indexed.
Locked in audioEven when the answer existed, finding it meant remembering which episode it lived in. That mental index didn't scale past a handful of memorable moments, and newcomers met hundreds of titles with nowhere to start.
No way inThe assistant had to be in front of the community in time for their Scale Up Cruise, not after a half-year build. That set the shape of everything: prove it on a slice, ship a thin surface, expand in stages.
Weeks, not quartersA demo that works once but costs a fortune per month at archive scale isn't a solution. The constraint was as much financial as technical, from day one.
Community budgetTranscription, retrieval, and grounded generation turn the same hours of audio into something a founder can ask a question of, and get a cited answer back in seconds.
A small, managed AWS footprint the team can run without a dedicated platform engineer. Infrastructure as code, and budget alerts from day one.
The assistant ships as an embeddable widget inside Swisspreneur's own site, in its icon and its brand. Nobody has to go somewhere new to ask a question. It opens with the promise stated up front and the questions people ask most, so a first-time visitor has somewhere to start instead of a blank input. One embed, one retrieval pipeline: answer quality is owned in one place.
Every answer carries its source: the quote, the guest, the episode number, the exact minute, and links straight to Spotify or Apple. When the answer spans the archive, the assistant draws from several episodes at once and cites each one. That's the difference between a generic chatbot and a research tool.
Managed retrieval, cheap vector storage at archive scale, managed transcription. Deliberately boring on the operational side, so the budget stays predictable and nobody needs a dedicated ML platform engineer. Per-query cost decides whether the community uses the assistant freely or rations it, so it was designed in rather than measured after. A typical answer lands around $0.02.
A founder's question hits the archive and comes back with the specific episode that answers it, not a wall of results.
Instead of scrolling hundreds of titles, a visitor starts from something they care about and follows it into the catalog.
Transcription rolled out in stages keeps the running cost where it needs to be, so the assistant stays on for everyone rather than rationed.
Picked for cost-at-scale and operability after handoff: managed inference, retrieval that holds up at archive size, and a substrate the team can run without a dedicated ML platform engineer.
They didn't sell us on the magic. They built a one-hour proof of concept that showed us exactly what was possible and where the trade-offs were. That's why we trusted them with the archive.
AI features for community products either feel native or get ignored. The choice wasn't about who could build the cleverest model, but who could ship something the community would reach for.
The value sat inside long-form audio, which makes this a transcription and retrieval problem, not a search bar added to a website. And the economics are part of the engineering: the stack holds up at archive scale, not just in a demo.
The value sat inside long-form audio. That meant transcription, semantic retrieval, and grounded generation — not a search bar bolted onto a website. Streaver builds audio-to-knowledge systems and could speak to the trade-offs without hand-waving.
A community product has to fit existing habits, not impose new ones. The assistant had to feel native to how Swisspreneur's audience already moves through the site — a drop-in widget, not a chatbot parachuted in from somewhere else. Product design was scoped in from the first conversation.
A managed AWS stack picked specifically because the cost profile holds up at archive scale. The financial design of the stack is part of the engineering, not an afterthought once the demo works.
Streaver works as a partner, not a pair of hands. The team was in it from day one — bringing ideas to the table, helping make the calls, and finding the option that actually fit.
Held the thread between Swisspreneur and Streaver, and doubled as a second brain on the engineering: architecture and direction.
Drove the initial discovery and design — shaping the widget's interaction and how a cited answer is presented, so the assistant feels native to Swisspreneur's brand rather than bolted on.
Core engineering across the build: the ingestion pipeline (episode audio + metadata → Bedrock Data Automation → knowledge base) and wiring the widget to grounded, cited answers.
Copilot and second brain on the engineering effort — the embeddable widget and its states, and iterating on the UI and answer quality.
First paying enterprise customers in sixteen weeks
Building a $1M product for $125K with a non-technical CEO at the keyboard
Podcast networks, associations, internal training libraries — anyone holding hours of audio or video nobody can search. Done thoughtfully, AI turns it into a community feature, not a chatbot bolt-on. Talk to us about a Knowledge Hub for your archive.