|
I'm Nataraj Angappan β a Cisco ISE and Wireless Technical Lead at Cognizant, working across enterprise network security, NAC, routing and switching, and Catalyst wireless. I build AI tooling that makes that work faster and more explainable: RAG assistants that explain authentication flows, agent workflows that keep a human approving every consequential change, and a small language model that diagnoses Wi-Fi RF problems across 2.4, 5 and 6 GHz. My bias is toward grounded answers, evidence trails and vendor-neutral design over flashy demos. |
name: Nataraj Angappan
role: Cisco ISE & Wireless Tech Lead
company: Cognizant
location: Coimbatore, India
focus: NAC Β· Wireless Β· RF Β· AI Automation
languages: Python Β· TypeScript
building: RF Root Cause SLM (QLoRA)
ISE Chatbot, Agentic-AI, BGP Automation.
repos: 42 public |
|
|
| π§ Languages & Core |
|
| βοΈ Backend & Frontend |
|
| π€ AI & Automation |
|
| π Networking & Security |
|
| Project | What it does |
|---|---|
| πΆ SLM β Wi-Fi RF issues | Diagnoses 2.4 / 5 / 6 GHz RF problems and returns a root cause with a supporting evidence chain. Vendor-neutral schema, QLoRA fine-tune, RAG for regulatory facts, FastAPI + React, and an ESP32 hardware probe. |
| π ISE Flow Assistant | A RAG chatbot that explains Cisco ISE and RADIUS/AAA flows β and draws them. |
| π°οΈ Catalyst 9800 DR Readiness | Scores wireless disaster-recovery readiness from controller CLI state and plays an operator-gated site failover, on pyATS/unicon mocks. |
| π§ͺ netsim lab | Simulates Cisco ISE / NAC, wireless and Catalyst Center scenarios on an 8 GB laptop, no CML required. |
| πΈοΈ Knowledge Graph RAG | Measures how much required evidence Graph RAG surfaces versus vector-only RAG on multi-hop questions. |
| π§° IT Ops Assistant | Multi-agent IT-operations command center with human-in-the-loop manager approval. |
- Ground it or don't say it. Numbers and policy claims come from retrieval with a citation, not from a model's memory.
- Show the evidence. An answer without a trail isn't usable in operations.
- Keep a human in the loop for anything that changes a production network.
- Vendor-neutral where it counts β normalise at the boundary instead of retraining or rewriting.

