Nucleus One
Permission-aware retrieval inside a multi-module enterprise platform
LangChain · LangGraph · ASP.NET Core 8 · SQL Server · React · TypeScript · Azure AI Foundry
The problem
Employees on Nucleus One — Acme One’s multi-module enterprise platform, built on a unified SQL Server schema of roughly 100 tables — needed to ask open-ended questions in natural language and get answers drawn from live platform data. The obvious risk: a chatbot with broad data access can quietly become a path around the platform’s own permission model, surfacing records a user was never supposed to see.
The approach
A Retrieval-Augmented Generation chatbot engineered and deployed into Nucleus One’s production environment, owned end to end from prototype through go-live. Retrieval grounding was iterated through prompt engineering and tuning of chunking strategy and embedding configuration, measured against answer accuracy on internal documents. The same platform work includes sole ownership of the Project-One module (Epic → Feature → Story → Task tracking, sprint planning, Kanban boards, analytics dashboards, delivered from SQL Server schema through ASP.NET Core 8 services to the React/TypeScript interface), the HR-One service layer and schema, and the platform’s JWT/RBAC authentication controller — plus ongoing work building LangChain and LangGraph agentic pipelines with tool-calling agents, stateful graphs and human-in-the-loop approval steps against the same APIs and data layer.
The interesting decision
Permission-aware retrieval. Every request is resolved against the asking user’s existing authorisation in the application before anything is retrieved, so the chatbot can only surface records that user is already entitled to see through the platform’s normal access controls — it cannot become a side channel around them. This is the single most interview-ready detail in the whole body of work: it’s a security decision expressed as a retrieval architecture, not a prompt-engineering trick.
What I’d do differently
Retrieval grounding today is validated through prompt iteration and manual accuracy checks against internal documents — useful, but informal. Building a structured, repeatable evaluation harness around it, in the spirit of the one now wired into Project-Ease, would turn "iterated until it looked right" into a measured, comparable baseline the same way.