Fed Aura Capital & Lenora
The platform you’re joining
Fed Aura Capital’s AI platform didn’t arrive fully formed. It was built around four transformation pillars, each solving a specific production problem that the prototype couldn’t handle. Understanding those pillars is the fastest way to understand why the platform looks the way it does today.
The four transformation pillars
Pillar 1 - Fast, Flexible, and Efficient Inferencing
"From token consumer to token producer"
Fed Aura Capital used to pay per-token to an external LLM vendor. Every API call left the network, costs were unpredictable, and there was no control over which model version served a request.
To keep data inside the network, make costs more predictable, and control model versions, the platform now runs self-hosted models with vLLM, uses llm-d for distributed inference, and exposes them through a managed API gateway via Models as a Service (MaaS).
| Before | After |
|---|---|
Per-token billing to external vendors |
Fixed infrastructure cost, self-hosted models |
Data leaves the network on every call |
Data stays in-cluster, full sovereignty |
Every team deploys its own model copy |
Single deployment shared via MaaS gateway |
No usage metering or access control |
Subscription tiers with API key management |
Related Sections: 2.1 Model Catalog & Deployment, 2.2 Deploy a Model, 2.3 Deploy a Model with llm-d, 2.4 Models as a Service
Pillar 2 - Connecting Data to Models and Agents
"From generic to useful AI"
A deployed model is still generic. Lenora could not quote Fed Aura Capital’s rates, did not know internal policies, and gave textbook answers with disclaimers.
The bank experimented with ways to make Lenora a more reliable domain expert. It refined her instructions using system prompts stored in the Prompt Registry and connected her to private bank documents so she could ground her answers in trusted information. Using AutoRAG, the bank tested different ways to retrieve and use those documents to find the configuration that worked best for Fed Aura Capital.
| Before | After |
|---|---|
Generic system prompt ("You are a helpful AI assistant") |
Domain-specific persona configured for mortgage lending |
Hallucinated policy answers |
RAG-grounded responses citing internal documents |
No way to update model knowledge without retraining |
Upload new documents; knowledge updates in minutes |
Related Sections: 3.1 System Prompts, 3.2 RAG & Knowledge Grounding, 3.3 AutoRAG
Pillar 3 - Accelerate Agentic AI Innovation
"From useful to productive AI"
Lenora could answer domain-specific questions, but could not take action. When a borrower asked "How much house can I afford?", she estimated a range with assumptions. An agentic Lenora calls the affordability calculator and returns a precise figure with a debt-to-income (DTI) breakdown.
The platform connects agents to real tools through the MCP Ecosystem, captures every decision path in MLflow traces, and continuously measures quality with automated evaluation pipelines.
| Before | After |
|---|---|
Regular chatbot: estimates with disclaimers |
Agentic chatbot: precise results from real tool calls |
Black-box agent decisions |
Full MLflow traces showing every tool call and LLM invocation |
Manual quality spot-checking |
Automated evaluation pipelines with EvalHub |
Related Sections: 4.1 Agentic AI & Tracing, 4.2 Tools and MCP, 4.3 Continuous Evaluation
Pillar 4 - Securing AI from Sandbox to Production
"From prototype to production-grade"
Agents that take real actions in a regulated industry must be secured before production. The CISO blocked deployment until a proper security model was in place. Regulators required evidence that the AI cannot be manipulated and maintains a full audit trail.
The platform now deploys NeMo Guardrails at the inference boundary, uses Garak for adversarial testing, and runs agents inside OpenShell sandboxes with deny-by-default policies and full audit logging.
| Before | After |
|---|---|
No input/output filtering |
NeMo Guardrails at the inference boundary |
Untested adversarial robustness |
Garak adversarial test results available for review |
Agents deployed as ad-hoc containers |
OpenShell with deny-by-default policies and OCSF ( Open Cybersecurity Schema Framework) audit logging |
No security observability |
Structured audit events, SIEM-ready |
Related Sections: 5.1 AI Safety, 5.2 Guardrails & Adversarial Testing, 5.3 OpenShell Agent Sandboxing
The personas
Throughout the workshop, you’ll interact with the platform as different Fed Aura Capital employees. Each persona has different access levels and sees different behavior from Lenora - this is intentional, and understanding why is part of the lab.
| Persona | Name | Role | What you’ll observe |
|---|---|---|---|
Borrower |
Sarah Mitchell |
Customer applying for a mortgage |
Asks about rates, gets policy-grounded answers, sees RAG in action |
Loan Officer |
James Torres |
Reviews borrower applications, manages pipeline |
Sends prompts that trigger tool calls, checks MLflow traces |
CEO |
David Park |
Executive oversight and analytics |
Queries produce analytics responses grounded in internal data |
Engineer |
You |
Supporting the AI platform |
Observes all of the above, checks traces, explores guardrail configs |