Biscuit AI
From conversational prototype to configurable platform.
A shopper asks half a question, the catalogue changed this morning, and the answer still has to be right.
Controlled recommendation, live catalogue retrieval, evaluation, speed and memory — on kiosks that serve shoppers all day, with no staff.
Product discovery had to resolve incomplete intent, retrieve current catalogue evidence, compare real attributes, preserve context, and decline to recommend when evidence was insufficient.
Nixense led the architecture and a five-engineer team from a working RAG prototype to a controlled, configurable product engine.
A supervisor resolved intent before routing to bounded specialists. Every operation could return insufficient evidence. Recommendation passed through a required-attributes gate grounded in live catalogue data.
Nixense separated the reasoning engine from configurable personas, product attributes, tool permissions, and category rules. A second retail vertical ran on the same engine through configuration, with no core-logic changes.
Representative answers fell from 3–4 seconds to 1.5 seconds. The work combined asynchronous execution with integration and testing of the response path, supporting kiosks operating around the clock.
Nixense built the team’s first RAG evaluation framework, separating answer relevance, context relevance, and groundedness. Regression checks could identify the failing layer instead of collapsing every defect into one score.
Session state retained current turns, criteria, and follow-up reference. Durable preference used a structured summary read on return, connected only through explicit distillation after interaction.
Biscuit AI publicly presents a physical-AI platform across kiosks, robots, displays, and applications. This case is limited to Nixense’s April 2024–March 2025 engagement and does not imply Nixense built every capability or deployment presented today.
The assistant improved when conversation stopped being treated as one model response and became a controlled path through intent, evidence, state, tools, validation, and operation.