Building AI Systems That Ship
>>> AGENTIC SYSTEMS · ARCHITECTURE TO PRODUCTION TRAFFIC
I take agentic systems from architecture to production traffic. Four years of multi-agent orchestration, RAG, and MCP tooling across retail, e-commerce, voice, and vision surfaces - built for extensibility, so new capability ships by configuration rather than redeployment.
Career
Four roles at Nagarro since 2022. The work itself is below.
| REC | ROLE | LEVEL | PERIOD | DUR |
|---|---|---|---|---|
| 04 | Associate Staff AI Engineer Own the architecture of a multi-tenant agent platform, and lead engineering enablement on AI in the SDLC.
| Staff | Jul 2026 – Present | 1M |
| 03 | Senior AI Engineer Led AI delivery end to end - document intelligence, computer vision, process automation - and took on RBAC and data governance for SOC 2.
| Senior | Feb 2025 – Jul 2026 | 1Y 5M |
| 02 | AI Engineer Built the first production agent systems on Google ADK and Vertex AI, across chat, voice, and document surfaces.
| Mid | Jul 2023 – Feb 2025 | 1Y 7M |
| 01 | Associate Engineer Real-time platform engineering - streaming, WebSockets, and microservices at concurrent scale.
| Junior | Feb 2022 – Jul 2023 | 1Y 5M |
Selected Work
Seven systems shipped to production. Each one names the role it was built in.
Three-tier agent orchestration - a central planner, seven domain subagents, and parallel ReAct workers on LangGraph - serving web chat, WhatsApp, and voice.
Workers run in parallel, so added capability does not accumulate serial latency. Zero-code multitenancy through a four-level configuration hierarchy (Global → Brand → Region → Persona) onboards new brands and regions without a deployment, with a model-agnostic LiteLLM gateway routing per tenant. Backend endpoints become agent tools automatically from their OpenAPI specs, annotated with workflow-chaining metadata so the planner can compose multi-step flows against endpoints it has never seen. Persistent semantic memory with a consolidation layer carries personalisation across sessions, a 14-tab admin console puts prompt, model, and flow control in non-engineers’ hands, and distributed tracing attributes latency and token cost per agent node.
7 domain subagents · parallel workers · new brands ship by config alone
Conversational analytics over a live enterprise backend, where the agent discovers its own tools from an OpenAPI 3.1 spec.
RS256/JWKS JWT auth in front of a FastAPI service on Cloud Run Gen2, routing queries through a Gemini orchestrator that composes multi-step API workflows autonomously. Explicit prompt caching with stale-cache auto-recovery cuts token cost, backed by asynchronous Firestore session persistence and fire-and-forget push notifications with exponential-backoff retry and a dead-letter fallback.
Zero-code tool discovery - new endpoints become agent capabilities with no code change
Bilingual Thai/English commerce assistant on LINE, with citation-backed answers across product, promotions, and FAQ.
A single-agent / multi-tool architecture on Google ADK selects between five specialised search tools by query intent. Hybrid relevance scoring combines semantic similarity, keyword match, and position rank with dynamic drop-factor filtering; Vertex AI Discovery Engine supplies citations. Fed by a knowledge-base pipeline handling PDF, DOCX, XLSX, PPTX, HTML, JSON and image ingestion through GCS resumable uploads, with incremental sync and zero-downtime orphan cleanup.
5 intent-routed tools · 3 search domains · citation-backed answers in two languages
Invoice intake and categorisation at 90%+ accuracy, with a conversational layer over the resulting expense data.
AWS Textract paired with custom models reaches 90%+ automated categorisation and sub-second responses across 10,000+ monthly transactions, cutting manual processing by 70%. A LangGraph assistant answers natural-language expense queries with persistent context, alongside a dashboard for trend analysis and predictive forecasting.
90%+ accuracy · 10,000+ transactions/month · 70% less manual effort
Two intake channels - speech and scanned documents - feeding one order pipeline for a Fortune 500 healthcare client.
Hot-word detection and multi-level variant selection let customers place complex orders by voice alone. An AWS Textract path handles scanned and photographed purchase orders through the same resolution logic, so both channels share line-item and variant handling. Processes 10,000+ orders monthly.
10,000+ orders/month · Fortune 500 client · one pipeline, two modalities
Custom facial recognition reading multiple skin parameters in real time, with an explainable recommendation engine.
92% detection accuracy across diverse skin types and conditions, processing multiple parameters simultaneously at 200ms latency. The recommendation engine surfaces its reasoning rather than returning an opaque score, so users see why a product was suggested.
92% detection accuracy · 200ms latency · 5,000+ users
Real-time video streaming fused with e-commerce - 1,000+ concurrent viewers buying inside the stream at sub-second latency.
AWS Interactive Video Services for the stream, WebSockets for live chat and product showcase, with dynamic pricing and one-click purchase inline. An admin console handles stream management, analytics, and multi-language support across web, mobile, and OBS, backed by REST APIs and socket connections spanning 50+ microservices.
1,000+ concurrent users · sub-second latency · 50+ microservices
Skills
- Expert
- Advanced
- Intermediate
- AI/ML & Agents
- Multi-Agent Systems
- LangGraph
- LangChain
- Google ADK
- RAG
- Context Engineering
- Function Calling
- MCP Servers
- Agent Skills
- Agent Memory
- LiteLLM
- Knowledge Graphs
- Conversational & Voice
- Conversational AI
- Voice Agents
- Speech-to-Text
- Intent Routing
- Multi-Turn Memory
- Multilingual NLU
- GenAI & LLMs
- Claude API
- Gemini API
- OpenAI SDK
- Structured Outputs
- Streaming
- Prompt Caching
- Guardrails
- HF Transformers
- AI Systems Engineering
- LLM Observability
- Distributed Tracing
- Token Cost Optimisation
- Model Routing
- Multi-Tenant AI
- RBAC & Governance
- Computer Vision
- OpenCV
- DLib
- OCR (Textract)
- PyTorch
- TensorFlow
- scikit-learn
- Visual Search
- Backend
- FastAPI
- Python
- REST
- Node.js
- Express.js
- WebSockets
- Microservices
- Event-Driven
- Frontend
- React.js
- Next.js
- TypeScript
- Tailwind CSS
- Cloud
- GCP
- AWS
- Vertex AI
- Cloud Run
- Lambda
- AWS IVS
- Data & Storage
- Qdrant
- Vertex AI Search
- Hybrid Search
- Reranking
- PostgreSQL
- Firestore
- SQLAlchemy
- DynamoDB
- DevOps & Tools
- Docker
- GitHub Actions
- Camunda BPM
- Terraform
- Git
- Linux
- AI-Assisted Development
- Claude Code
- GitHub Copilot
- Custom MCP
- Agent Tooling
About
Most AI demos don’t survive production. Mine do.
I own the full lifecycle - architecture, implementation, deployment, governance, and the 3am incident that proves it works. Open to senior engineering roles, freelance projects, and AI consulting.
- Multi-Agent Orchestration
- Planner, subagent, and parallel worker hierarchies on LangGraph and Google ADK - systems that decompose a request, act on it, and stay traceable while doing so
- Conversational & Voice AI
- Assistants across chat, WhatsApp, LINE, and speech - intent routing, tool use, multi-turn memory, and hot-word voice ordering at 10,000+ orders a month
- RAG & Context Engineering
- Hybrid retrieval, reranking, and citation-backed answers, with prompt caching and context budgeting that keep quality up and cost down
- AI Observability & Governance
- Distributed tracing for per-node latency and token cost, output-safety guardrails, RBAC and data governance - the layer that makes an agent system operable
- Extraordinary Role Model
- Top 1% of Nagarro employees · 4+ consecutive years for exceeding performance expectations and delivering high-impact AI solutions
- 2022–2026
- AWS R&D Event Presenter
- Represented Nagarro's innovation in AI/ML and real-time commerce systems at the AWS R&D event
- 2023
AI in the SDLC: sessions & workshops
I run internal sessions and workshops on AI-assisted development: agent configuration, prompt and context strategy, custom tooling, and folding Claude Code and Copilot into day-to-day engineering work.
- Bachelor of Technology, Information Technology
- Acropolis Institute of Technology and Research
- 2018–2022
Get in Touch
Open to senior engineering roles, freelance projects, and AI consulting.
- Within 24h
Prefer a casual chat? Let's discuss your project over a quick call.
Schedule Coffee Chat →- Tanmay Jain is an Associate Staff AI Engineer at Nagarro Software, promoted into the role in July 2026 after serving as Senior AI Engineer. He owns the architecture of a multi-tenant agent platform and leads engineering enablement on AI in the SDLC.
- LangGraph, LangChain, and Google ADK for multi-agent orchestration; FastAPI for backend services; OpenCV and DLib for computer vision. On the model side: Claude API (Anthropic SDK), Gemini API, and OpenAI SDK, routed through LiteLLM for provider-agnostic fallback.
- He architected a production three-tier hierarchy - a central planner, seven domain subagents, and parallel ReAct workers on LangGraph - serving web chat, WhatsApp, and voice for a multi-tenant commerce platform. Earlier he built single-agent, multi-tool architectures on Google ADK with intent-routed tool selection.
- Yes. Bilingual Thai/English assistants on LINE with citation-backed retrieval, conversational analytics over live enterprise APIs, a natural-language expense assistant, and voice-enabled ordering with hot-word detection and multi-level variant selection processing 10,000+ orders monthly for a Fortune 500 healthcare client.
- Distributed tracing across agent workflows gives per-node latency and token cost attribution, so a slow or expensive step is identifiable rather than inferred. Cost control comes from prompt caching that separates large cacheable instruction from small per-request context, plus model routing and fallback through a LiteLLM gateway.
- Three shapes. A short-term project (typically 2 to 8 weeks) for a feature build, MVP, or proof-of-concept. An ongoing retainer of roughly 10 to 20 hours a week for part-time AI engineering. Or a one-time architecture review: a deep-dive into an existing AI stack, delivered as written recommendations. Scope and cost are set on a short call before anything starts.
- Yes, and most engagements start that way rather than greenfield. Typical entry points are adding an LLM or agent layer to a product that already exists, replacing a prototype that does not hold up in production, or reviewing an AI stack that works but is slow, expensive, or hard to change. His day job is a platform other engineers build on, so integrating with someone else's conventions is the normal case.
- Yes, remote-first and globally. He is based in Indore, India on IST (UTC+5:30), which gives a full working overlap with Europe and the Middle East, morning overlap with APAC, and an early-morning-to-midday overlap with US East Coast. Written handover is the default so progress does not depend on a live call.
- The problem in plain language, one example of a request and the answer you would want back, whatever data or API surface the system has to work against, and the constraint that actually matters most - latency, cost per request, accuracy, or data residency. Architecture follows from the binding constraint, so naming it early avoids building the wrong thing well.
- Yes - open to senior engineering roles, freelance projects, and AI consulting engagements. Based in Indore, India (IST, UTC+5:30), working remotely with clients globally. Response within 24 hours at tanmayjain717@gmail.com.