Open to AI/ML & Software Engineering roles

Govind Mudavadkar

AI/ML Engineer·Software Engineer

I build and deploy agentic AI from proof of concept to production, with the evaluation, observability and guardrails that keep it reliable once it's live.

Get in touch See the work
Govind Mudavadkar
01 / About

Demos are easy. Production is the job.

Most GenAI work dies somewhere between proof of concept and production, and the gap is rarely the model. It's everything around it. Retrieval quality you can measure instead of eyeball. Hallucination rates you track instead of hope about. Guardrails, cost and latency budgets, traces you can actually debug from. Closing that gap is what I do at Pfizer, where I build LangGraph agents, multimodal RAG and MCP tooling, plus the evaluation and observability that keep them honest.

I'm a software engineer as much as an ML one. I came up through backend work (Java, Spring Boot and REST APIs at BMC Software), trained PyTorch vision models as a research assistant, and now build full-stack products end to end: FastAPI services, React front ends, Docker and Kubernetes on AWS. Alongside it I write research, with seven-plus peer-reviewed papers across medical imaging, document intelligence and multi-agent systems.

Role
AI/ML Engineer
Company
Pfizer Inc
Based in
Boston, MA / New London, CT
Degree
M.S. Computer Engineering
School
Northeastern University
GPA
3.8 / 4.0
Focus
Agentic AI · ML · SWE · Evals · Observability
Status
Open to opportunities
02 / Experience

Four years across AI research, platform engineering and backend systems.

AI/ML Engineer

Jan 2026 - Present
Pfizer Inc·Groton, CT
LangGraphMCPAWS Bedrock FastAPIReact / TSKubernetes RAGASLangSmithLiteLLM Observability

AI/ML & Software Automation Engineer, Co-Op

Jan 2025 - Sep 2025
Pfizer Inc·Groton, CT
RAGKnowledge GraphsHuggingFace Spring BootApache CamelYOLO Bayesian OptDataikuPostgreSQL

Research Assistant, Data Science & Software

Feb 2024 - Dec 2024
Northeastern University·Boston, MA
PyTorchONNX / INT8LibTorch (C++) SageMakerDistributed GPUSpark Elasticsearch

Software Development Engineer

Jan 2022 - Jul 2022
BMC Software Inc·Pune, India
JavaSpring BootVue.js PostgreSQLMavenJUnit
03 / Education

Computer engineering, with a research habit.

Northeastern University

M.S., Computer Engineering
Boston, MA Sep 2023 - Dec 2025 GPA 3.8 / 4.0

University of Pune

B.E., Computer Science & Engineering
Pune, India Jun 2019 - May 2023 GPA 3.8 / 4.0
04 / Skills

The stack I reach for.

Grouped by what I use it for, not by how impressive the logo looks.

Generative & Agentic AI

LLMs (Claude, GPT, Llama)Generative AI Agentic AILangGraphLangChain MCPRAGPrompt Engineering GuardrailsEvaluation (LangSmith, RAGAS) Langfuse

Model Training & Serving

Fine-tuning (LoRA / PEFT)Post-training (SFT, GRPO) RL from Verifiable RewardsPyTorch TensorFlowHuggingFace Transformers vLLMSelf-hosted LLM Serving Inference OptimizationConstrained Decoding Quantization (NF4, INT8, fp8)ONNX

ML, Vision & Modelling

NLPComputer Vision (YOLO, Faster R-CNN) ResNet / ViTVLMsXGBoost EnsemblesBayesian Optimisation Statistics

Languages

PythonJavaTypeScript JavaScriptC++C# SQLRScala CUDABash / Linux

Backend & Web

FastAPISpring BootREST APIs MicroservicesNode.jsReact Next.jsAngularTailwind MVCApache Camel Event-Driven Architecture

Data & Storage

PostgreSQLMySQLMongoDB RedisElasticsearchNeo4j PineconeMilvusChromaDB SparkKafkaMapReduce Knowledge GraphsETL Pipelines

Cloud, DevOps & MLOps

AWS BedrockS3SageMaker LambdaEC2DynamoDB Azure AI FoundryGCPDocker KubernetesGitHub ActionsJenkins TerraformCI/CDMLflow Observability & Monitoring Dataiku DSSHPC / Slurm

Practices

System DesignData Structures & Algorithms OOPDesign PatternsAgile / Scrum LLM EvaluationA/B Testing Benchmark DesignStatistical Testing Reproducibility Testing (JUnit, PyTest) Technical Writing
05 / Projects

Things I built to find out whether they'd work.

Research prototypes, GPU systems and full-stack products. Most are open source.

LedgerMind

LLM post-training · Verified inference
  • A 4B open model, LoRA-fine-tuned then trained with GRPO against a deterministic verifier, answers FinQA questions statistically tied with a frontier model (78.5% vs 80.3%, McNemar p = 0.38) at roughly 600× lower cost per query when self-hosted.
  • The model never writes the final number: it cites figures with source pointers and a computation plan, a safe AST executor runs the plan, and a verifier checks every figure against the document, so answers arrive with a full audit trail or a reason for refusal. Served on vLLM at 48.7 req/s in fp8 on a single H200.
GRPOLoRA SFTvLLM PyTorchfp8 / bf16FinQA SlurmPython

Resilient Multi-Agent ESG Auditor

Multi-agent RAG · Research
  • Four cooperating LangGraph agents (Supervisor, Retriever, Auditor and Critic) audit SEC 10-K filings for ESG disclosures, running fully local on Llama 3.1/3.2 via Ollama.
  • A coverage-gap-aware Critic using three-sample self-consistency cut refusal rate 48% → 20% and raised substantive answers 52% → 76% on a tier-stratified 25-question evaluation.
LangGraphLlama 3.1Hybrid RRF BM25RAGASPython

Deal Intelligence AI

Hierarchical Tree-Graph RAG
  • Parses 300-page M&A and credit agreements from SEC EDGAR into unified hierarchical JSON trees, then indexes them across a Neo4j knowledge graph and a LlamaIndex vector baseline.
  • Supports clause extraction with section-level citations, cross-document conflict detection and multi-document synthesis, exactly where flat vector RAG loses document structure.
LlamaIndexNeo4jLangGraph PageIndexDoclingPython

GPU K-mer Counter

CUDA · High-performance computing
  • A two-stage CUDA pipeline hashes billions of overlapping genomic k-mers on-GPU and scatters them into disjoint partitions, so every kernel's working set stays inside device memory.
  • Six documented optimisation iterations (v1 → v6.2) with per-partition hash tables and stream compaction The whole performance path is preserved in the repo.
CUDA 11C++17GPU Memory Stream CompactionMake

Melanoma Detection at Scale

Distributed deep learning
  • Multi-GPU PyTorch DistributedDataParallel training across 4× Tesla K80s with all-reduce gradient synchronisation on a 12.2 GB dermoscopic image dataset.
  • Dask-based CPU parallelism for image enhancement with measured speedup and efficiency curves across 1 to 8 cores, run on Northeastern's Discovery HPC cluster.
PyTorch DDPCUDADask Transfer LearningHPC

Automated Cloud Deployment Pipeline

Infrastructure-as-Code · DevOps
  • Architected a cloud-native user onboarding system on AWS using SNS, Lambda and SendGrid behind secure VPC boundaries, security groups and explicit ingress rules.
  • Automated CI/CD with EC2 Auto Scaling, ELB and GitHub Actions running tests, baking AMIs and performing rolling updates, reducing downtime by 35%.
AWSLambdaSNS EC2 Auto ScalingGitHub ActionsIaC

Talent Palette

Full-stack · MERN
  • A responsive MERN platform for artists built with React, TypeScript and object-oriented design, enabling job creation and personalised content feeds.
  • Scalable REST APIs with JWT auth improved content discovery by 20%; deployed on AWS EC2 with Kubernetes and Jenkins.
ReactTypeScriptNode.js MongoDBJWTKubernetes Jenkins

Present: A Better Way to Present

Full-stack · Next.js
  • A presentation platform built around one idea: create a deck and share it with nothing but a link. No exports, no attachments, no account friction for viewers.
  • Slide management with transitions, themes, import/export and live editing, built on Next.js API routes and MongoDB, deployed to Vercel.
TypeScriptNext.jsNode.js MongoDBVercel

Blog Generation with LLaMA 2

Open-source LLM · Applied
  • An end-to-end blog generation app running Meta's open-source LLaMA 2 locally, orchestrated with LangChain and served through a Streamlit interface.
  • Written as a step-by-step build guide covering prompt templating, model loading and quantisation trade-offs for running a 7B model on consumer hardware.
LLaMA 2LangChainStreamlit Python
All repositories on GitHub
06 / Publications

Peer-reviewed research.

Multimodal document intelligence, applied ML platforms and medical imaging.

KnovaAI: Advanced Multimodal AI for Chemistry-Aware Document Intelligence

AIChE2026

Machine Learning-Driven API Development via Automated Platforms

AIChE2025 Read

Gastric Cancer Detection with Ensemble Deep Learning

MDPI Diagnostics2024 Read

4+ additional peer-reviewed papers

Google Scholar Full list
Google Scholar profile
08 / Connect

Let's build something measurable.

Fastest way to reach me is email. Everything else is below.

Email mudavadkar.govind@gmail.com LinkedIn govind-mudavadkar GitHub @Govindrm7 Google Scholar Publications Medium @govindmudavadkar

Hiring for AI/ML or software engineering?

I'm open to roles where research and production engineering sit in the same job description. Happy to walk through any system on this page.

Email me LinkedIn GitHub Google Scholar Medium