AI — Vignesh Athiappan

In production

The enterprise assistant — multi-agent, on Microsoft Teams, org-wide

WhatA central router classifying every ask and dispatching to specialist production agents — leave, resource information, dev-ops sync, knowledge — grounded by RAG over country- and domain-specific policy indexes with a query-rewriting layer. Multilingual by design. Demoed to the CEO and COO.

Agent securityLeast-privilege tool access per agent and an auditable action registry — every capability explicit, scoped, and logged.

Write-ups: from a chatbot to a real agentic assistant → · a Teams bot on Azure, end to end →

Data agents on the lakehouse — ask the warehouse a question

WhatA dynamic GraphQL layer exposed as MCP tools: the model reads the schema, constructs its own queries, and answers conversationally. Data access for operations, finance, HR and leadership — demoed to the CEO.

Write-up: turning a system of record into an AI agent →

Agents for talent acquisition — requisitions, worked by agents

WhatPurpose-built agents supporting recruitment requisition workflows — drafting, matching and moving them along, with humans approving.

Models in day-to-day use — Azure AI Foundry, per application

WhatFoundry model deployments serving everyday features across applications — summarization, translation, classification, insight — chosen and cost-engineered per workload.

Grounding

MCP servers — systems of record as safe AI tools

WhatMultiple MCP servers built and deployed, wrapping enterprise systems as tools a model can reason over — scoped, logged, and permission-aware.

Knowledge graph — an ontology for a warehouse with no foreign keys

WhatBuilt from the code layer — table relationships, per-column definitions, join paths — so schema context stops being hand-fed into every prompt. The grounding layer between the warehouse and the agents.

Write-up: a knowledge graph with no foreign keys →

Foundry IQ & Fabric IQ — the grounding roadmap

WhatMicrosoft's semantic-grounding direction maps almost one-to-one onto the knowledge-graph design — evaluated first-hand at Microsoft's AI Innovation Day and folded into the roadmap.

Write-up: field notes from AI Innovation Day →

In the engineering loop

AI reviewing the code — a PR review desk, benchmarked

WhatAn LLM-powered pull-request review desk wired into Azure DevOps — then benchmarked against commercial reviewers before the team committed to it.

Write-ups: building the review desk → · benchmarking it →

AI triage, zero backend — one file, one queue, one model

WhatA single-file dashboard reading Azure DevOps directly and using a model to triage the support queue — classification, summaries, suggested owners — with no infrastructure to own.

Write-up: the zero-backend triage dashboard →

Currently sharpening — model fine-tuning on Azure

NowWorking the fine-tuning ladder — SFT and LoRA on Azure — from baseline evals toward trained adapters.

Not experiments. Systems in daily use.

Hiring for this?