Challenge
Knowledge lived across wikis and tickets; employees needed trustworthy answers without leaking sensitive content.
Case study · 2025
AI, ML & Generative AI
Generative AI copilots grounded in your data
Generative AILLM IntegrationNestJSPythonApplication Security
Client
Enterprise knowledge & support org
Engagement
LLM copilot with RAG
Duration
8 months
Year
2025
Lumen Copilot grounded generative answers in enterprise knowledge without leaking sensitive content. Retrieval pipelines, prompt policies, eval harnesses, and a NestJS gateway with authZ and audit logs make every interaction attributable. Tool-calling is guarded; citations are first-class.
An LLM-powered assistant with RAG over enterprise knowledge, NestJS orchestration, and guarded tool-calling.
Narrative
How the engagement moved from constraints to a shippable system.
Approach
We built retrieval pipelines, prompt policies, eval harnesses, and a NestJS gateway with authZ and audit logs.
Solution
Lumen Copilot answers with citations, supports tool actions, and keeps every prompt/response auditable.
Scope
What we delivered
RAG pipelines over wikis, tickets, and docs
NestJS gateway with authZ and audit logs
Prompt policies and safety filters
Eval harness with citation scoring
Tool-calling framework with allowlists
Support and employee chat experiences
Highlights
What stood out
45% reduction in tier-1 support tickets
Citation accuracy above 92% on eval set
Role-based knowledge scopes enforced
Full prompt/response auditability
Engagement flow
A repeatable rhythm from discovery through hardening — tuned to this product’s constraints.
01
Knowledge & risk map
Classified corpora, access boundaries, and high-risk answer classes.
02
Retrieval & gateway
Built indexing, retrieval, and the NestJS orchestration layer.
03
Eval & policy
Stood up eval harnesses, prompt policies, and red-team scenarios.
04
Rollout
Phased seat expansion with audit reviews and support metrics.
Technology stack
The delivery stack behind this engagement — client surfaces paired with services, data, and infrastructure.
Frontend
React / NuxtChat UICitation Panels
Backend
NestJSPythonRAG PipelineTool-calling Gateway
AI & Security
LLM IntegrationGenerative AIApplication SecurityEval Harness
Outcomes
01
45% reduction in tier-1 support tickets
02
Citation accuracy above 92% on eval set
03
Role-based knowledge scopes enforced
Looking ahead
Expand tool actions into ticketing systems with stronger human-in-the-loop gates.
More work
Onboard the Apeiriion express
Tell us about your product, platform, or modernization goal — we'll assemble a senior squad and a clear delivery plan.