An enterprise copilot that survived legal and security review
A B2B SaaS vendor needed RAG-based product assistance grounded in customer docs - not a generic chatbot.
The challenge
Product leadership wanted AI-assisted onboarding fast, but security required tenant isolation, citation of sources, and clear guardrails before any beta.
How we approached it
Data and risk audit first
Audited document types, PII patterns, and retrieval quality before choosing an architecture.
Tenant-scoped retrieval
Built tenant-scoped vector indexes with citation-backed answers inside the existing product shell.
Evaluation before beta
Added evaluation harnesses for hallucination rate, latency, and escalation to human support.
Studio integration
Partnered with Agam Studio for API integration, feature flags, and staged rollout controls.
Outcomes
Security-approved beta
Beta launched with legal-approved data handling and per-tenant access boundaries.
Support load reduced
Support ticket volume for onboarding questions fell in pilot accounts.
Right sequencing
Pilots were sequenced against real data readiness - avoiding a costly misfire.
Technologies involved
- Python
- LangChain
- OpenAI
- PostgreSQL
- pgvector
- React
- AWS
In their words
“They helped us sequence AI pilots against real data readiness - saved us from a costly misfire.”