# Dheeraj Bhavsar: AI Engineer and Engineering Manager

> AI engineer and engineering manager in Bengaluru. Builds agentic AI, LLM, MCP and knowledge-graph systems and leads 15+ engineers shipping enterprise AI for Fortune 500 companies.

Website: https://www.dheerajbhavsar.com
Email: dheerajbhavsar940@gmail.com
Location: Bengaluru (Bangalore), Karnataka, India
Profiles: https://x.com/dheeraj_bhavsar | https://www.linkedin.com/in/bhavsardheeraj | https://github.com/BhavsarDheeraj | https://stackoverflow.com/users/7688783/dheeraj-bhavsar

## Bio

Dheeraj Bhavsar is an AI engineer and engineering manager based in Bengaluru (Bangalore), India. He is an Engineering Manager at QuesttAI, where he leads 15+ engineers and owns enterprise AI delivery end to end: technical pre-sales, solution architecture, BOM and TCO estimation, client cloud readiness, build, deployment and hypercare. He is still hands-on: he is the primary author of QuesttAI's enterprise context-layer platform, a knowledge graph of business entities, metrics and decisions served to AI agents over the Model Context Protocol (MCP), shipped to Nokia, Vodafone Idea, PepsiCo, ZEE5 and AM/NS India. He also owns the company's DevSecOps and compliance layer (CI/CD, container hardening, SSO/IAM, SOC 2 and GDPR readiness). He has 6+ years of experience across agentic AI, LLM applications, backend systems, cloud and mobile.

## Roles he fits

AI Engineer, Engineering Manager, Agentic AI Engineer, Generative AI (GenAI) Engineer, LLM Engineer, AI Solutions Architect, Forward Deployed Engineer, Player-coach engineering leader

## Enterprise clients

Nokia, Vodafone Idea, PepsiCo, ZEE5, AM/NS India (ArcelorMittal Nippon Steel), Motilal Oswal, House of Anita Dongre, LivSpace

## Experience

### Engineering Manager, QuesttAI, Bengaluru (Sep 2025 - now)

- Lead 15+ engineers and own enterprise AI delivery end to end: technical pre-sales, solution architecture, BOM and cost estimation, client infra readiness, build, deployment and hypercare.
- Own the company's DevSecOps and compliance layer: CI/CD, container hardening, secrets and access management, InfoSec reviews for global enterprises, SOC 2 and GDPR readiness.
- Guide client cloud and security teams on landing zones, networking, IAM/SSO and egress; deploy on AWS and Azure inside client tenants with no client-data egress.
- Still hands-on: primary author of the core context-layer platform and its licensing and secure-distribution model, shipped to Nokia, Vodafone Idea, PepsiCo, ZEE5 and AM/NS India, often with PwC India.
- Pre-sales across 20+ enterprise accounts; a reworked LLM mix cut one client's projected 5-year TCO from ~$1.0M to ~$0.6M. Built the hiring loop (40+ interviews) and performance process.

### Software Development Engineer II, QuesttAI, Bengaluru (Oct 2021 - Sep 2025)

- Built the backend of a GenAI tutor (FastAPI, tool-calling agents, real-time WebSockets), then diagnosed its RAG limits and led the move to knowledge-graph context that became the company's core platform.
- Led Motilal Oswal's email-automation bot from data mapping and MSSQL-to-Postgres migration to production (Dec 2025), with 90%+ resolution/escalation accuracy in evaluation.
- Architected a retail BI and conversational-analytics platform for House of Anita Dongre.
- Migrated a PHP monolith to Node.js microservices, saving ₹5L+/month; built a WebRTC live-classes platform driving ₹1Cr+/month revenue.

### Senior Software Engineer, MobCast, Mumbai (Jul - Oct 2021)

- Developed an API caching layer reducing response times by 70%.
- Engineered native iOS and Flutter bridging; introduced silent push notifications to remove polling.

### Software Development Engineer, Squareboat, Gurugram (Jul 2020 - Jul 2021)

- Delivered apps across ed-tech, health-tech, cyber-security, fin-tech and travel, used by 10M+ people.
- Implemented SAML auth, SSL pinning and MITM detection; built CI/CD with Jenkins and Fastlane.

## Education

B.E. Computer Science and Engineering, Rajiv Gandhi Technical University (RGPV), Bhopal, 2016–2020

## Skills and stack

- AI: LangGraph, LangChain, MCP, OpenAI and Claude APIs, RAG, knowledge graphs, multi-agent systems, guardrails
- Backend: Python, TypeScript, FastAPI, Node.js, NestJS, microservices
- Data: PostgreSQL, DuckDB, MongoDB, Redis
- Cloud & security: AWS, Azure, GCP, Docker, Kubernetes, Terraform, Entra ID, SOC 2, GDPR
- Observability: LangSmith, Datadog, Grafana, Sentry, CloudWatch
- Realtime & mobile: WebRTC, WebSockets, Asterisk, iOS (Swift), Flutter, React, Next.js

Areas of expertise: Artificial Intelligence, AI Engineering, Agentic AI, AI Agents, Multi-Agent Systems, Generative AI, Large Language Models, LLM Evaluation, Model Context Protocol (MCP), LangGraph, LangChain, Retrieval-Augmented Generation (RAG), Knowledge Graphs, Context Engineering, Decision Intelligence, Engineering Management, Solution Architecture, Technical Pre-sales, DevSecOps, SOC 2 and GDPR compliance, AWS, Microsoft Azure, Kubernetes, Docker, Python, TypeScript, FastAPI, Node.js, PostgreSQL, DuckDB, WebRTC, iOS (Swift), Flutter

## Case studies

### Enterprise Context Layer

URL: https://www.dheerajbhavsar.com/projects/intelligence-warehouse
Client: Core platform | Role: Architect & EM | Year: 2024-Present | Stack: Python, PostgreSQL, FastAPI, LangGraph, MCP, DuckDB, Next.js

The context layer for enterprise AI - entities, metrics and decisions in one traversable graph, so agents follow the same decision paths a company's best people do.

Problem: Enterprise AI pilots fail on context, not models. Decision logic lives in people's heads, spreadsheets and ERPs, so agents built on raw tables guess - and nobody can audit why.

Approach:
01. Modelled the business as a three-layer graph - entities → metrics → decisions - where every node is a versioned, validated Python function with provenance and lineage.
02. Built the builder studio consultants use on-site: scope, data-graph harvesting, workflow capture → codegen → validate → simulate.
03. Kept routing deterministic: agents walk the graph and run validated node code, never ad-hoc SQL, so every answer is traceable.
04. Served the graph to any agent framework over MCP - model-agnostic, with every traversal logged.

Impact:
- Live with enterprises including Nokia, Vodafone Idea, ZEE5, AM/NS India and PepsiCo, largely delivered with PwC India.
- Fortune 100 telecom: failing AI pilot to 96% decision accuracy in 6 weeks (published case study).
- Each additional use case in ~11 days at under 10% of the first one's cost.

Metrics: 96% decision accuracy (case study); 6 wks kickoff → production; 11 days per next use case

### Secure AI Appliance & Licensing

URL: https://www.dheerajbhavsar.com/projects/iw-appliance
Client: Core platform | Role: Architect & Owner | Year: 2026 | Stack: Docker, Nuitka, PostgreSQL, AWS, Azure, Entra SSO, Caddy

Turned the context-layer platform into a licensed, encrypted, self-installing appliance that runs inside the client's own cloud - no client data ever leaves their tenant.

Problem: Enterprise clients refused hosted SaaS - their data could not cross the public internet - but shipping the product into their clouds put the IP and the licence terms at risk.

Approach:
01. Packaged the platform and its agent as single-tenant VMs (Docker, Postgres 16, SSO, HTTPS-only egress) with a guided installer and a `doctor` diagnostic.
02. Built a control plane: enrollment tokens, leases, appliance heartbeat, machine-bound activation, two-person kill and release attestation.
03. Compiled and encrypted backend images with Nuitka and shipped them through a private registry with per-client pull scopes.
04. Wrote the deployment canon - reference infra, security & governance, client deployment guide, MCP reference, BOMs - and ran the AWS and Azure rollouts.

Impact:
- Deployed inside client clouds at Nokia, Vodafone Idea, AM/NS India and Multiples Equity.
- Zero client-data egress: the client owns infra and IAM; we ship images and docs.
- Started with a contract clause, hardened into full licensing once the deal sizes justified it.

Metrics: 0 client-data egress; AWS+Azure client-cloud installs; 1-cmd guided install

### Expert Knowledge-Capture Agent

URL: https://www.dheerajbhavsar.com/projects/morrie
Client: Core platform | Role: Engineer → Architect | Year: 2024-Present | Stack: LangGraph, LLMs, MCP, FastAPI, WebSockets

Started as a GenAI tutor; became the agent that interviews domain experts, captures their decision logic into the context layer, and answers questions by walking the graph.

Problem: Capturing how experts actually decide took months of interviews and hand-modelling - and our first RAG-based tutor showed that retrieval alone cannot reason over a business.

Approach:
01. Built the 2024 tutor backend: FastAPI, tool-calling agents and real-time WebSockets.
02. Diagnosed its RAG failure modes and moved context into a knowledge graph - the seed of the context-layer platform.
03. Re-cast the agent as a conversational interviewer that turns expert answers into graph nodes, rules and thresholds.
04. Shipped it as a web app and an MCP server, so client agents query the same source of truth.

Impact:
- Expert knowledge captured by conversation instead of questionnaires.
- 69 procurement workflows captured across Fiber, Marketing and Scrap procurement at Vodafone Idea.
- One agent interface for both building and querying the context layer.

Metrics: 69 workflows captured (one POC); NL expert interface; MCP served to any agent

### Distributor Order Engine

URL: https://www.dheerajbhavsar.com/projects/order-engine
Client: PepsiCo · with PwC India | Role: Technical Lead | Year: 2026 | Stack: Python, FastAPI, PostgreSQL, AWS SQS, WebSockets, Next.js

AI-assisted distributor ordering for PepsiCo (AIDEN) - from application architecture and engineering standards to pilot hypercare.

Problem: Distributor orders arrived through messy, inconsistent channels; recommendations had to land in front of ~3k users reliably and inside strict enterprise security.

Approach:
01. Wrote the application architecture and engineering standards, signed off by the client.
02. Designed the event path ML → SQS → backend → SSE/WebSockets, S3 presigned uploads and a BFF behind Okta/Apigee.
03. Ran a 57-field order-line data-quality audit that reset the data contract with the client.
04. Instituted change-request control and led the client-facing delivery.

Impact:
- Reached pilot hypercare in Oct 2026.
- Data gaps surfaced early: 41 of 57 fields flagged, ~43% cell completeness.
- Standards adopted as the baseline for the build.

Metrics: Pilot hypercare reached; ~3k users notified; 57-field data audit

### Supply Chain AI Agents

URL: https://www.dheerajbhavsar.com/projects/supply-chain-agents
Client: FMCG Enterprise | Role: Lead Engineer | Year: 2024-2025 | Stack: Python, LangGraph, AI Agents, FastAPI, MongoDB

Purpose-built agents for inventory allocation, demand sensing and distribution optimization in real time.

Problem: Balancing service levels against working capital across warehouses, distributors and outlets is a constant, high-dimensional trade-off no dashboard can solve.

Approach:
01. Built agents for inventory allocation, demand sensing and distribution optimization.
02. Coordinated decisions across warehouses, distributors and retail outlets.
03. Tied every recommendation to real-time operational signals.
04. Grounded the agents in a shared context layer for unified decisions.

Impact:
- Real-time allocation balancing service levels and working capital.
- Coordinated multi-echelon decisions instead of siloed ones.
- Field-ready recommendations, not just reports.

Metrics: Real-time allocation; Multi-echelon coordination; FMCG scale

### Sales Intelligence Platform

URL: https://www.dheerajbhavsar.com/projects/sales-intelligence
Client: Retail Enterprise | Role: Lead Engineer | Year: 2024-2025 | Stack: Python, AI Agents, FastAPI, PostgreSQL, React

AI-driven sales optimization - beat planning, outlet scoring and rep guidance tied to real-time signals.

Problem: Field sales decisions - which outlet, which SKU, which day - were made on gut feel, disconnected from inventory and market reality.

Approach:
01. Built beat planning and outlet scoring driven by real-time signals.
02. Generated rep guidance factoring inventory positions, competition and trends.
03. Surfaced decisions at the field level, not the boardroom.
04. Integrated with a shared context layer for unified decisions.

Impact:
- Field-level decisions driven by live signals at scale.
- Rep guidance grounded in inventory and market dynamics.
- Higher-quality beat plans and outlet prioritization.

Metrics: Field-level decisions; Real-time signals; Scale rep guidance

### AI Retail Analytics Platform

URL: https://www.dheerajbhavsar.com/projects/retail-analytics
Client: House of Anita Dongre | Role: Architect | Year: 2025-2026 | Stack: Python, PostgreSQL, FastAPI, LangGraph, Next.js

Retail BI and conversational analytics - a warehouse schema, stock-history modelling and an agent for store- and SKU-level decisions.

Problem: Decision-makers needed store- and SKU-level answers, but data was scattered across sources with no conversational way to reach it.

Approach:
01. Designed the retail data-warehouse schema and integrated multi-source data.
02. Proposed stock-history storage and indexing to keep SKU-level queries fast.
03. Captured the brand's business ontology so the agent speaks its language.
04. Shipped conversational analytics and KPI dashboards over the warehouse.

Impact:
- Conversational access to store- and SKU-level intelligence.
- Unified multi-source data into one queryable warehouse.
- Faster, self-serve decisions for merchandising teams.

Metrics: Store+SKU granularity; Conversational analytics; 1 unified warehouse

### Email Support Bot

URL: https://www.dheerajbhavsar.com/projects/email-support-bot
Client: Motilal Oswal Financial Services | Role: Project Lead | Year: 2025 | Stack: Python, FastAPI, PostgreSQL, LLMs, Langfuse, AWS EKS

Automated classification, resolution and escalation of customer emails - led from data mapping to production in Dec 2025.

Problem: Motilal Oswal's support team faced a high, growing volume of customer email, with slow and inconsistent first responses.

Approach:
01. Mapped the CRM fields and migrated the history from MSSQL to Postgres, de-duplicating along the way.
02. Built LLM classification with clustering to route resolution vs escalation.
03. Sized and deployed on EKS with autoscaling; added Langfuse observability and sentiment detection.
04. Ran post-go-live triage and redesigned the classifier from production feedback.

Impact:
- 90%+ resolution and escalation accuracy in evaluation.
- Went live in production in Dec 2025.
- Freed the support team for complex, high-value cases.

Metrics: 90%+ eval accuracy; ~5,000 email eval corpus; Dec 2025 live in production

### Procurement Audit Pipeline

URL: https://www.dheerajbhavsar.com/projects/procurement-audit
Client: LivSpace | Role: SDE II | Year: 2022-2023 | Stack: OCR, Python, FastAPI, PostgreSQL

OCR-driven three-way reconciliation (PO - Invoice - GRN) that automated manual procurement auditing.

Problem: Procurement audits were fully manual - matching POs, invoices and goods-receipt notes by hand, slow and error-prone.

Approach:
01. Integrated CRM data with OCR models for document extraction.
02. Automated three-way reconciliation of PO, Invoice and GRN.
03. Flagged mismatches and exceptions automatically.
04. Moved auditors from matching documents to reviewing exceptions.

Impact:
- Automated the manual matching step of procurement audits.
- Three-way reconciliation end to end.
- Faster, more accurate procurement close.

Metrics: 3-way reconciliation; Auto document matching; OCR extraction

## FAQ

### Who is Dheeraj Bhavsar?

Dheeraj Bhavsar is an AI engineer and engineering manager based in Bengaluru (Bangalore), India. He is an Engineering Manager at QuesttAI, where he leads 15+ engineers and owns enterprise AI delivery end to end: technical pre-sales, solution architecture, BOM and TCO estimation, client cloud readiness, build, deployment and hypercare. He is still hands-on: he is the primary author of QuesttAI's enterprise context-layer platform, a knowledge graph of business entities, metrics and decisions served to AI agents over the Model Context Protocol (MCP), shipped to Nokia, Vodafone Idea, PepsiCo, ZEE5 and AM/NS India. He also owns the company's DevSecOps and compliance layer (CI/CD, container hardening, SSO/IAM, SOC 2 and GDPR readiness). He has 6+ years of experience across agentic AI, LLM applications, backend systems, cloud and mobile.

### Is Dheeraj Bhavsar an AI engineer?

Yes. Dheeraj Bhavsar is a hands-on AI engineer and engineering manager. He builds agentic AI systems with LangGraph, LangChain and the Model Context Protocol (MCP), designs knowledge-graph context layers for LLM agents, and deploys them inside enterprise AWS and Azure tenants. He wrote most of QuesttAI's context-layer platform himself while leading a team of 15+ engineers.

### What has Dheeraj Bhavsar built?

An enterprise context layer (a knowledge graph served to agents over MCP) that took a Fortune 100 telecom from a failing AI pilot to 96% decision accuracy in 6 weeks; a secure, licensed AI appliance that runs inside client clouds with zero data egress; an email support bot for Motilal Oswal with 90%+ evaluation accuracy, live since Dec 2025; a distributor order engine for PepsiCo with PwC India; supply-chain and sales-intelligence AI agents; and a retail analytics agent for House of Anita Dongre.

### What is Dheeraj Bhavsar's tech stack?

Python, TypeScript, FastAPI, Node.js and NestJS; LangGraph, LangChain, MCP, OpenAI and Anthropic Claude APIs, RAG and knowledge graphs; PostgreSQL, DuckDB, MongoDB and Redis; AWS, Azure, GCP, Docker and Kubernetes; plus earlier work in iOS (Swift), Flutter and WebRTC.

### What kind of roles does Dheeraj Bhavsar fit?

Player-coach engineering leadership: Engineering Manager, Head of Engineering, AI Engineering Lead, Agentic AI Lead, AI Solutions Architect or Forward Deployed Engineering Lead, especially at product companies and startups building AI for enterprises.

### Where is Dheeraj Bhavsar based?

Bengaluru (Bangalore), Karnataka, India.

### How do I contact Dheeraj Bhavsar?

Email dheerajbhavsar940@gmail.com, or reach him on LinkedIn (linkedin.com/in/bhavsardheeraj) or X (@dheeraj_bhavsar).