What I Built
Vector Chat is an AI client-intelligence platform I built to help professionals retain and use the messy, high-signal information that lives across meetings, notes, and documents—so they can walk into every client interaction prepared, with answers grounded in source material.
Building it end-to-end as a production-ready, multi-tenant SaaS (RAG assistant over a client knowledge base) taught me what it actually takes to build production-ready AI for knowledge-intensive workflows:
What I Learned
- •Trust beats novelty: In knowledge work, adoption hinges on verifiability. Designing citation-backed responses made it clear that explainability and traceability are product requirements, not "nice-to-haves."
- •Data quality is the moat: The hardest problems weren't the chat UI—they were ingestion, organization, and retrieval quality. The project reinforced that AI outputs are only as reliable as the system that curates context.
- •Product thinking under constraints: I had to make explicit tradeoffs across latency, cost, and accuracy, and design guardrails (e.g., context management) so the system stays useful as complexity scales.
- •Operational realities matter: Multi-tenant isolation and subscription/usage-based billing forced a disciplined view of risk, permissioning, and unit economics—how features translate into sustainable operations.
- •Workflow-first design: "Meeting intelligence" features (capture → summarize → follow-up) taught me to start from real user workflows and incentives, then decide where AI genuinely reduces cognitive load.
How Do I Evaluate New Tech?
Going forward, this project is the reference point for how I evaluate new tech: map capability to a concrete workflow, identify where trust can break, quantify the operational/economic constraints, and then decide what's actually deployable versus what's just impressive in a demo.