RetailGPT — Agentic Retail Intelligence Platform
Ask retail data a question in plain English and get a governed, cited, multi-format answer back — safely.
Role: Lead / solution architect — concept-to-production, designed and built end to end.
The problem
Retail teams sit on large transactional datasets but can’t query them without an analyst and a wait. The hard part of a natural-language layer over that data isn’t the chat UI — it’s the security and governance around it: who can ask what, what’s safe to execute directly, and what needs retrieval instead.
Business Context
Retail teams at Datacue sat on large transactional datasets they couldn't query without an analyst and a wait. The goal wasn't just a chat UI over that data — it was making a natural-language layer safe enough to trust with real business questions.
Architecture
Key capabilities
Dual-path orchestration — deterministic warehouse queries and non-deterministic RAG retrieval run in parallel and merge into one response
AI security layer for jailbreak and prompt-injection defense, plus per-user category access control
Multi-format responses generated in parallel — summary, table, chart and export
A recommendation engine that reads the same governed data to suggest next actions
Automatic primary/secondary LLM fallback so the assistant stays available
Full conversation history and AI observability from day one
Security & Risk
The real risk in a natural-language layer over transactional data isn't the chat UI — it's authorization: who can ask what, what the agent may execute directly versus only retrieve, and what happens when a query touches data outside the requester's permission. RetailGPT enforces that boundary before generation, not after.
Outcome
A governed, cited, multi-format answer back to the business user — safe by construction, not by review after the fact.
Challenges
Running deterministic warehouse queries and non-deterministic RAG retrieval in parallel, then merging them into one coherent response, was the core architectural challenge — on top of enforcing per-user category access control and jailbreak/prompt-injection defense before generation happened at all, not as a review step afterward. Keeping the assistant available also meant building automatic primary/secondary LLM fallback into the same request path.
Lessons Learned
Authorization has to be enforced before generation, not filtered out of the answer afterward. Deciding what a user may ask, what the agent may execute directly versus only retrieve, and what data is out of scope has to happen at the query boundary — a “safe” answer generated first and checked after is just a policy violation nobody has noticed yet.
Architecture pillars
Designed and reviewed against all ten — see how I evaluate every architecture.
Want this level of review applied to your own system? Start an AI Architecture Review.
Technology
Want something like this, built properly?
I design and ship systems like this one — from architecture through to production.