Omnichannel multi-agent platforms
Supervisor, Planner and Worker agents on LangGraph, reaching users on WhatsApp, web, mobile and social over chat, voice and email.
Hyderabad, India · Co-founder, Agni AI Labs
Kumara Vathadi
Senior AI Engineer | Agentic AI Architect
Enterprise-grade omnichannel multi-agent AI, built end to end.
- 6 years in AI/ML
- Architect across banking, real estate, retail & maritime
- IIT Madras M.Tech
About
Building AI products from architecture to production
I'm a Senior AI Engineer with 6 years of experience, working as an architect across multiple projects and building products end to end — from requirements and architecture through build and Azure/Kubernetes deployment to observability and production rollout. I specialise in enterprise-grade, omnichannel, multi-agent AI for real estate, banking, retail and maritime, reaching users on WhatsApp, web, mobile app, Facebook and Instagram over chat, voice and email. I also bring quantitative ML depth in Bayesian price elasticity and brand equity for CPG. As co-founder of Agni AI Labs, I'm building the Defragment AI platform end to end for restricted domains.
What I build
Four things I do well
RAG & Text2SQL agents
Hybrid retrieval with BM25 and cross-encoder reranking on Azure AI Search, plus Text2SQL agents over enterprise data via scalable MCP.
LLM observability & guardrails
Every node traced with Langfuse, Arize AX and Phoenix for token cost, latency and debugging; RBAC, content safety and audit trails built in.
Bayesian price-elasticity & brand-equity ML
Hierarchical Bayesian models for CPG pricing, promotion and market mix, deployed on the cloud with automated retraining.
Selected work
Projects
Defragment AI — Metadata-Driven Multi-Agent Platform
Problem One scalable platform that serves restricted domains end to end and extends domain by domain without rebuilding the core.
- Co-founded Agni AI Labs and building Defragment AI end to end as a product.
- A scalable, metadata-driven, multi-domain, multi-agent AI platform — covering architecture, build and testing.
- Multi-agent
- Metadata-driven
- Multi-domain
COMPASS AI — Omnichannel Multi-Agent Platform
Problem One conversational platform for real-estate sales operations — broker and agency registration, commission distribution, bookings, EOI, support tickets and roadshow approvals — with property search and full tracking for BD, HOD and Leadership.
- Omnichannel solution on WhatsApp and a mobile app serving external personas (agencies, brokers) and internal personas (BD, HOD, Leadership).
- Layered architecture: orchestration with Redis/PostgreSQL memory, guardrails and RBAC → Query Analyzer, Planner and Response Synthesizer supervisor agents → business and reasoning agents over an A2A service bus.
- Proactive, event-driven agents and a provider-agnostic LLM layer (LangChain / LiteLLM), with every LangGraph node traced in Arize AX.
- LangGraph
- Azure AI Foundry
- Azure Service Bus
- Databricks
- Azure AI Search
- PostgreSQL
- Redis
- Content Safety
- AKS
- APIM
- WhatsApp Cloud API
Omnichannel Agentic AI Platform
Problem One consistent, compliant customer experience across phone, web chat, email and social channels for banking queries, authentication and fraud-related requests.
- Stateful LangGraph graph with confidence gating and dynamic routing across Greeting, Authentication, Banking, Fraud Detection and RAG agents under a Supervisor, with a compliance layer and full audit logging.
- Real-time voice with tool calling on Azure AI Voice Live and Azure Communication Services telephony, with automatic language detection (English, Malay, Mandarin).
- Hybrid BM25 + cross-encoder RAG on Azure AI Search with multilingual prompt templates, served from an async FastAPI backend on AKS.
- LangGraph
- Azure AI Voice Live
- Azure AI Search
- Azure Communication Services
- FastAPI
- WebSockets
- Jinja2
- AKS
Real-Time Analytical Insights Chatbot
Problem Lets analytics teams query enterprise sales, panel and ratings data conversationally instead of writing SQL or waiting for reports.
- Collaborative multi-agent platform of SQL, Insights, Chart and Planner agents, with A2A context passing for multi-step, cross-KPI reasoning.
- Scalable MCP for stateless connectivity to Azure ADLS and PostgreSQL, deployed on AKS and evaluated with Langfuse.
- LangGraph
- Scalable MCP
- A2A
- Azure ADLS
- PostgreSQL
- AKS
- Langfuse
Maritime Compliance & Communication AI Platform
Problem Prioritising high-volume fleet email and tracking regulatory compliance (CII, FuelEU) across a global fleet.
- Email management on Microsoft Graph API that prioritises, categorises and summarises maritime communications, with AI-drafted responses from specialised agents.
- Real-time CII and FuelEU compliance modules computing annual status, risk probabilities and forecasts, with scenario-simulation agents for recommendations.
- Microsoft Graph API
- AKS
- PostgreSQL
- Simulation AI
- Specialised Agents
Price Elasticity & Promotion Optimization
Problem Quantify price sensitivity and coupon effectiveness to guide pricing, promotion and spend-allocation decisions.
- Hierarchical Bayesian and Lasso elasticity models for 7,000+ products across 40+ retailers and 2 e-commerce platforms.
- Parallelised modelling with Spark UDFs on Databricks, cutting modelling time by 80%.
- Price, promotion and new-product launch optimisation frameworks deployed in an AKS-hosted planning tool.
- Hierarchical Bayesian
- Lasso / Ridge
- Log-Log Regression
- Databricks
- Spark UDFs
- Azure ADF
- AKS
Brand Equity & Market Mix Modelling
Problem Estimate brand equity and its long-term effect on sales and advertising returns.
- Bronze → Silver → Gold → Harmonized pipeline combining social, survey and retail POS data, with brand equity estimated via Confirmatory Factor Analysis.
- Brand equity integrated into Market Mix Modelling to quantify long-term advertising effects.
- Near real-time pricing dashboards (3-hour refresh) on Databricks and Azure Data Factory.
- CFA
- Bayesian Modelling
- Random Forest
- Databricks
- Azure ADF
- Mixed-Effects Models
Tech stack
Tools I work with
Agentic AI
- LangGraph
- Scalable MCP
- A2A (Azure Service Bus)
- Text2SQL
- RAG (hybrid, BM25, reranking)
- pgvector
- Jinja2 prompts
- Guardrails
- Tool calling
- LangChain
- LiteLLM
- Azure AI Foundry
- Azure OpenAI
LLM observability
- Langfuse
- Arize AX
- Arize Phoenix
- OpenInference
- OpenTelemetry
Azure & cloud
- AKS
- Azure Front Door
- WAF
- APIM
- AI Search
- AI Foundry
- Voice Live
- Content Safety
- Service Bus
- Databricks
- PostgreSQL
- Redis
- ADLS
- Data Factory
- Docker
- Azure DevOps CI/CD
- MLflow
Channels
- WhatsApp Cloud API
- Facebook Messenger
- Web chat
- Mobile app
- Voice
- Microsoft Graph API
- REST
- WebSockets
- Multilingual NLP (English / Malay / Mandarin / Manglish)
ML & analytics
- Bayesian & Hierarchical Bayesian Regression
- Lasso / Ridge
- Random Forest
- Price Elasticity
- Market Mix Modelling
- CFA
- Mixed-Effects Models
- Demand Forecasting
- Conjoint Analysis
- Deep Learning (CNN, TensorFlow)
Languages
- Python (FastAPI, asyncio, SQLAlchemy)
- SQL
- PySpark
- R
Experience
Experience & education
Experience
Senior AI ML Engineer — Tiger Analytics
Oct 2020 – Present · Hyderabad, India
Architect and deliver omnichannel multi-agent AI platforms for enterprise clients in real estate, banking, retail and maritime end to end, and lead cross-functional teams across concurrent projects.
Co-founder — Agni AI Labs
Building the Defragment AI platform end to end for restricted domains.
Education
M.Tech, Computational Mechanics — IIT Madras
2018 – 2020 · CGPA 8.99/10
B.Tech, Mechanical Engineering — RGUKT Nuzvid
2013 – 2017 · CGPA 8.92/10
Agni AI Labs
Defragment AI
As co-founder of Agni AI Labs, I'm building Defragment AI — a scalable, metadata-driven, multi-domain, multi-agent AI platform that serves restricted domains end to end and extends domain by domain without rebuilding the core.
Visit agniailabs.inContact
Let's talk
Open to conversations about agentic AI architecture, multi-agent platforms and end-to-end AI product builds.
- kumara.vathadi@gmail.com
- +91-96666 05244
- Hyderabad, India
- linkedin.com/in/vksvarma
- github.com/vksvarma