We took the client from a Notion-doc idea to a paid B2B AI copilot - multi-tenant SaaS, evaluation harness, agentic workflow engine and a brand that gets the sales meeting.
Client name protected under NDA · engagement details unchanged
Months from idea to paid cohort
Golden-eval pass rate
Median agent response (seconds)
Per-tenant isolation tests passing
01 / Background
the client was a thesis: mid-market revenue teams spend more time updating tools than working in them, and a well-grounded copilot could absorb that work. The founders had the domain and the network - they needed a senior team to build the product, the AI engineering behind it, and the brand that would get them onto enterprise shortlists.
02 / Challenges
No engineering team yet - every architecture decision was still open, and the runway demanded a paid cohort within two quarters.
Early prototypes were demo-shiny and production-fragile - long agent runs drifted off-task and hallucinated record updates.
No way to know if a prompt or model change improved or regressed real workflows - releases were a vibe-check.
Enterprise buyers needed hard guarantees about prompt isolation, log retention and per-tenant policy controls.
A crowded copilot market - the website had to make the differentiated thesis instantly legible.
03 / Solutions
Stood up a multi-tenant TypeScript/React/Postgres stack with strict row-level security, per-tenant secrets and a clean audit trail.
Built a deterministic agent loop on top of MCP-style tool use, with explicit plan/critique/act steps, tool-call timeouts and structured fallbacks.
Golden-dataset eval suite, regression gates on every PR, and shadow-traffic comparison between candidate prompts/models - releases are scored, not vibes.
Per-tenant policy layer for tool exposure, PII redaction in prompts and logs, structured-output validation before any write-back into CRM.
Per-conversation traces, token-cost attribution per workflow and per customer, latency SLOs surfaced inside the product.
Identity, narrative website, demo experience and pricing page - built to convert technical buyers in the second meeting.
04 / Stack & apps
05 / Outcomes
the client went from idea to its first paying enterprise cohort inside two quarters.
Evaluation discipline pushed pass rate above 90% on the golden workflow set - and kept it there release over release.
Multi-tenant isolation passed enterprise security reviews on the first pass.
The narrative site has become the sales team's most-used demo asset.
Related work
Next case · 15
A founder-led skincare brand got the full treatment - identity, packaging system, art direction and a high-craft headless storefront that earns its premium pric…
