Tessera
An AI research workspace for biotech

Tessera's founders came to us with a hypothesis: biotech researchers waste 30% of their time on literature review that an AI workspace could compress into minutes. They had domain expertise and early customer conversations but no product and no engineering team.

We built the first working prototype in three weeks. The core was a RAG pipeline over PubMed and proprietary research corpora, with a workspace interface that let researchers ask questions in natural language, get cited answers, and save findings into structured research briefs. The first five beta users were onboarded before we'd finished the design system.

The AI architecture was Claude-first with a multi-provider abstraction. We used pgvector on Supabase for the embedding store, which let us avoid standing up a separate vector database for what was still an unproven product. The eval harness — which we built alongside the product, not after — tracked answer accuracy, citation quality, and hallucination rate across 200 gold-standard questions the founders curated with their beta users.

By month six, Tessera had 40 paying research teams and the metrics to raise. We helped them write the technical sections of the pitch deck, sat in on three investor meetings to answer architecture questions, and then spent the final three months hiring their founding engineering team and handing over every system we'd built.