# Anupam Kumar **Backend & Generative AI Engineer | Python · AWS · Azure · Kubernetes** +91 78763 02026 · anupamkumar.nith@gmail.com · Gurugram, India LinkedIn: linkedin.com/in/anupam8nith · GitHub: github.com/anupam8nith · Medium: medium.com/@anupam8nith · LeetCode: leetcode.com/u/anupam8nith --- ## Summary Backend engineer who builds Generative AI systems. Over the last couple of years I've taken LLM products from the first architecture call to production: designing the service, writing it in Python and FastAPI, shipping it on AWS and Kubernetes, then owning it once real users depend on it. Most of my work is at the application layer of LLMs. RAG pipelines, multi-agent chatbots, document-processing systems, and the retrieval, prompt design, and reliability work that keeps them dependable. I'm just as comfortable in the plumbing underneath: serverless on AWS, AKS and KEDA on Azure, and the CI/CD that keeps deploys boring. NIT Hamirpur graduate, happiest when a hard system-design problem meets a real deadline. --- ## Skills **Generative AI & LLMs:** LangChain, RAG (Retrieval-Augmented Generation), Multi-Agent Systems, AI Agents, Tool Use, AWS Bedrock, Azure OpenAI, Azure AI Foundry, Large Language Models (LLM), Prompt Engineering, Function Calling, Structured Outputs, Embeddings, Vector Search, Semantic Search, Model Context Protocol (MCP) **Languages:** Python, SQL, JavaScript, TypeScript, Bash, C++ **Backend & APIs:** FastAPI, Django, Flask, REST APIs, WebSocket, GraphQL, Server-Sent Events (SSE), async/await, asyncpg, Pydantic v2, OpenAPI **DevOps & Platform:** Docker, Kubernetes, CI/CD, GitHub Actions, Bitbucket Pipelines, CloudFormation, Kustomize, Git, SonarQube, pytest, mypy, flake8 **AWS:** Lambda, API Gateway, AppSync, DynamoDB, S3, SQS, SNS, EventBridge, Cognito, CloudWatch, IAM, Bedrock, Aurora, RDS, Serverless Framework, Boto3 **Microsoft Azure:** AKS, KEDA, AI Search, Azure OpenAI, Document Intelligence, AI Foundry, Service Bus, Event Grid, Blob Storage, Managed Identity **Databases:** PostgreSQL, DynamoDB, Qdrant (vector), Azure AI Search (vector), Redis **Architecture:** Serverless, Microservices, Event-Driven, Multi-Tenant SaaS, Distributed Systems, RAG Pipelines, Multi-Agent Orchestration, System Design, Agile / Scrum --- ## Professional Experience ### Q3 Technologies (Q3 Infotech Pvt Ltd) — Software Development Engineer **Oct 2023 – Present** | Gurugram, India Projects: Synthrics (AI bid-writing, Australia), Q3 Ask Contracts (RAG chatbot), Keypath Compass (Canvas LMS multi-agent), and BerthMaps and a UK rail delay-repay claims platform for Mistral Data. - Shipped the backbone of Synthrics' GenAI bid-writing platform: **three production microservices on Azure Kubernetes Service** with KEDA autoscaling, plus the **14-stage multi-tenant document pipeline** that runs on them. Picked the AKS and KEDA setup, and had it live in two months. - Own the backend and the architecture calls across **two production GenAI products** and a multi-agent client build, spanning RAG, vector search, and document AI. Where it mattered I chose between **AWS Bedrock** and **Azure OpenAI** on cost, latency, and ops fit, and built on LangChain, Qdrant, and Azure AI Search. - Cut our AWS iteration cycles **in half** and saved roughly **$9,000 a year** with a Docker-based Lambda debugger and selective invoker. It came out of the delay-repay and BerthMaps work, where redeploys were eating real time. - Took manual Lambda provisioning down about **75%** with a CI/CD pipeline that auto-deploys **124+ functions**, built on reusable CloudFormation modules, boto3, and file-hash change detection so only what actually changed gets shipped. - Turned new-project setup from a couple of hours into **under two minutes** with `setuptemplate`, a pip-installable scaffolder for FastAPI, Flask, and Django wired up with mypy, pytest, and flake8. The backend teams still use it. - Fixed **three critical production incidents** and wrote the post-mortems and handover docs the team still leans on. - Mentor three junior engineers on AWS (Lambda, API Gateway, DynamoDB, CloudFormation) and the GenAI stack (Qdrant, local LLMs, RAG), and run architecture reviews. Also ran an internal AWS VPC session for new joiners. ### Solid State Physics Laboratory (SSPL), DRDO — Research Intern **Jun 2022 – Aug 2022** | New Delhi, India - Studied **GaN HEMT** performance for telecom use with Silvaco simulations and TonyPlot, and co-authored a comparison against silicon benchmarks that fed into device-design decisions at SSPL. --- ## Key Projects ### AI Bid-Writing Platform — Document Indexing, Extraction & Summarisation Pipeline (Synthrics, Australia) **Feb 2026 – Present** · Python · FastAPI · Azure AKS · KEDA · Service Bus · Azure AI Search · Azure OpenAI · Document Intelligence · PostgreSQL (asyncpg) · Pydantic v2 · Docker · Kustomize - Owned the indexer, extraction, and summarisation consumers of a **14-stage multi-tenant pipeline**. Three AKS services, in production inside two months. - Killed the Azure OpenAI **429s under peak load** with token-aware embedding batching (100 texts / 15k tokens) and Retry-After backoff. Added a stale-job reaper, graceful SIGTERM handling, and a multi-tenant cache on Pydantic v2 and asyncpg. - Trimmed the Docker image **23%** (775 to 600 MB) and added a DOCX-to-PDF-to-Document-Intelligence fallback through headless LibreOffice for files Azure couldn't read directly. ### Q3 Ask Contracts AI — Production RAG Backend for Contracts Q&A **Jan 2026 – Feb 2026** · FastAPI · PostgreSQL · Qdrant · JWT · LangChain · AWS Bedrock · Ollama - Took the backend from an empty repo to production in **five weeks**: a JWT-secured RAG pipeline with SSE streaming, multi-page citations, async attachment indexing, and checksum dedup on uploads. - It's **live at Q3** and used daily by senior internal teams, who've been happy with the answer accuracy and citations. Locked it down on the way: pulled an SSO bypass, added CORS, rate limiting, security headers, and intent-based query routing. ### Keypath Compass Chatbot — Multi-Agent Canvas LMS Assistant (Azure AI Foundry) **Dec 2025 – Jan 2026** · Python · FastAPI · Azure AI Foundry · LangChain · Canvas LMS API - Delivered the first release. The client liked it enough to come back with a **follow-on proposal and a bigger scope**. - Designed the **4-agent setup** (intent classifier, internet-fallback, document agent, identity capability) for Canvas LMS Q&A. Refactored the `intentAgent` to clear every SonarQube issue across 7+ PRs, and kept the agents model-independent so swapping LLM versions doesn't break things. ### BerthMaps — Real-Time Train Berth Tracking Platform (Mistral Data) — [product page](https://www.mistral-data.com/product/berth-maps/) **Jul 2025 – Nov 2025** · Python · AWS Lambda · DynamoDB · AppSync · API Gateway WebSocket · S3 - Cut DynamoDB network overhead **25%** and Lambda runtime **20%** by reworking batch operations, and rebuilt the SVG upload path with S3 presigned URLs to get past the 10 MB cap. - Built the live-tracking layer on AppSync subscriptions over DynamoDB streams, so berth updates land in real time across regional rail. Cleared three production incidents along the way. ### Claims Management Portals — UK Rail Delay-Repay (Mistral Data) **Dec 2023 – mid-2025** · Python · AWS Lambda · Serverless Framework · CloudFormation · Aurora PostgreSQL · Loqate API - Built the delay-repay claims portals (customer and agent workflows) on AWS Lambda for a UK rail operator. Around **1,000 people use them daily at 99.5% uptime**. - Wired in Loqate for UK address and IBAN validation, shipped **38+ Lambda functions** through CI/CD, and cut claim processing time **40%** with conditional CSV and PDF reports off Aurora Postgres. --- ## Education **B.Tech, Computer Science & Engineering** — National Institute of Technology, Hamirpur **2019 – 2023** | CGPA 8.41 / 10.0 --- ## Achievements & Recognition - **Competitive programming:** Codeforces best rank **280**, CodeChef best rank **121**, and 900+ DSA problems across LeetCode, GeeksforGeeks, HackerRank, and Codeforces. - **Academic:** JEE Mains **99.08 percentile** (2019). - **Writing:** Wrote a Medium piece on Google ADK, multi-agent systems, and the Model Context Protocol (MCP); cross-posted to LinkedIn. - **Open source:** Building [HealAll India](https://healallindia.com), an open-source platform that helps people find medical assistance. A personal project, not tied to any employer ([github.com/HealAll-India](https://github.com/HealAll-India/HealAll)).