CHAPTER II / II.2
II.2
Source-grounded AI agent demo platform
AI demos are easy to make look convincing while hiding weak retrieval, invented claims, unreliable citations, and unsafe sharing. Pixelity needed a repeatable way to create focused company assistants from approved sources and make their quality inspectable before presenting them.
Context
AI demos are easy to make look convincing while hiding weak retrieval, invented claims, unreliable citations, and unsafe sharing. Pixelity needed a repeatable way to create focused company assistants from approved sources and make their quality inspectable before presenting them.
Build
Built an internal platform that discovers and reviews sources, creates a source-backed business brief, indexes approved content, and produces a citation-grounded assistant. Administrators can test the assistant internally and create an expiring, passcode-gated share link for a specific recipient.
Outcome
The system runs as Pixelity's internal demo engine, producing narrow, inspectable assistants for selected prospects from approved sources. Its conversational use remains internally tested; it is not an external production messaging system.
Key capabilities
Python/FastAPI backend with Next.js administration and guest surfaces.
PostgreSQL/pgvector indexing, with Redis and RQ background jobs.
Hybrid full-text and vector retrieval, Reciprocal Rank Fusion, optional reranking, MMR diversification, and same-source context expansion.
Citations reconstructed from authoritative stored metadata instead of generated freeform.
A deterministic, citation-free fallback for insufficient or high-risk evidence.
Golden-set evaluation across retrieval coverage, ranking quality, and absence handling.
Tenant isolation, audit trails, encrypted passcodes, guest-session isolation and expiry, and fail-closed rate limits.