BlackBear AI
Introduction
BlackBear AI was an AI consulting and product studio operating in the window right after ChatGPT arrived, when it was suddenly possible to build things that had been research projects a year earlier — and completely unclear which of them anyone would pay for.
BlackBear brought us in to be the technical side of that bet. They handled the market exploration and the commercial side; we owned the engineering. Three products shipped in under a year, each answering a different question.
Agent Lunar — a marketing department for companies that don’t have one
Agent Lunar is AI-first marketing software: a small business describes itself in a conversation, and the product generates a marketing strategy — positioning, target audience, content pillars, posting cadence, ninety-day goals — then writes the posts, generates the visuals, and schedules them across Instagram, Facebook and TikTok. CMO-as-a-service, effectively.
Agent Lunar’s content calendar: AI-generated posts scheduled per channel. Shown with example data.
The founding team behind it had the market insight but no technical capability — they came to BlackBear to have their product built. We took the technical lead: system architecture, the SvelteKit and Django stack, the strategy-generation pipeline, long-term memory built on Postgres with pgvector, and the multi-channel publishing integration.
It is a good illustration of what this kind of engagement is for. A non-technical founding team can get to a working, sellable product — but only if someone takes responsibility for the technical direction end to end, rather than handing over tickets.
Superhuman — one workspace instead of a drawer of tools
Superhuman was BlackBear’s own product. The thesis was that teams would want a single shared AI workspace rather than a growing collection of separate AI subscriptions — so we built a multi-tenant platform where a team installs the apps it needs: ChatGPT and DALL·E, chat-with-your-documents, a copywriting generator, SEO and research tools, each with per-team billing and spending limits.
How Superhuman was pitched: team access to the best models without per-seat subscription fees, with premade apps on top.
We led the engineering across the platform — the Django and SvelteKit application, the team and billing model, the app framework that let new tools be added without rebuilding the shell, and the document ingestion behind the knowledge-base apps.
Those document apps are the part that aged best: ingest a company’s own material, answer questions from it, cite the sources. It is now simply how everyone expects this to work, but at the time it had to be assembled from parts.
Charlotte — conversational job matching for Vivaldis
Charlotte is an AI recruitment assistant built for Vivaldis Interim, a Belgian staffing agency, and deployed on their public site. Candidates describe the work they are looking for in ordinary language, and Charlotte narrows down function, location, travel radius, contract type and language through conversation, then returns ranked, live vacancies alongside the chat.
Charlotte handling a job search in Dutch, with matched vacancies ranked in the panel beside the conversation. Example conversation and listings.
Two details make it more than a chatbot wrapper. The matching does not simply forward a query to the agency’s existing search engine: a second model pass derives descriptive tags from the conversation and re-ranks the raw results by semantic similarity — an early retrieval-augmented approach layered on top of a legacy enterprise matching API. And the assistant was deliberately fenced in: no health, politics, religion or financial topics, no competitor discussion, no collecting contact details mid-chat, with spontaneous applications routed by postcode to the correct regional office and a human recruiter.
Charlotte was already largely built when we joined the project, so this is a contribution rather than an authorship claim. It belongs here anyway, because conversational job matching is exactly the problem Talio productises today — the same idea, several years and a great deal more capability further on.
Results
- Three AI products designed, built and shipped in under a year, across marketing, workplace productivity and recruitment
- Technical lead on Agent Lunar and Superhuman; contributor on Charlotte
- Retrieval-augmented answering and re-ranking running in production well before it became standard practice
- Two full stacks — Django, SvelteKit, Postgres with pgvector, OpenAI, vector search and third-party publishing APIs — taken from empty repository to live product