Available for ML Engineering Roles

Atul Deshpande

ML Engineer | GenAI | Agentic Systems

Building production-ready AI systems with focus on retrieval-augmented generation, autonomous agents, and scalable machine learning infrastructure.

View Projects Get In Touch Download Resume
Atul Deshpande

About

I'm an ML engineer specializing in generative AI, agentic workflows, and retrieval systems. I build end-to-end solutions—from data pipelines and model training to deployment and evaluation—with an emphasis on reproducibility, scalability, and measurable impact.

My work spans multi-agent intelligence engines, RAG architectures for technical QA, computer vision applications, and competition-grade ML solutions. I'm passionate about turning research prototypes into reliable, production-ready systems.

Professional Experience

Machine Learning Engineer

Softron Technology

Jan 2026 – Jun 2026
  • Agentic AI Assistant: Developed an autonomous agent integrating real-time weather APIs, rule-based recommendation engines, and ML-driven fertilizer prediction workflows for agricultural decision support.
  • AI Pipeline Engineering: Built scalable inference pipelines using LangChain and Groq LLM APIs for contextual, multi-turn agricultural assistance with fallback handling and latency optimization.
  • Production Deployment: Completed industrial training in advanced ML architectures, computer vision, and NLP; successfully transitioned theoretical models into containerized, scalable deployment pipelines.
LangChain Groq API Agentic Workflows ML Pipelines Docker

Technical Product Intern

Inagiffy Solutions

Sep 2025 – Oct 2025
  • Cost Optimization: Reduced video production prototyping costs by 60% per unit by migrating proprietary API dependencies to self-hosted open-source models on dedicated GPU cloud infrastructure.
  • Multimodal Pipeline Design: Architected custom node-based workflows integrating voice cloning, image-to-video generation, and temporal consistency modules for automated synthetic media creation.
  • Infrastructure Strategy: Evaluated trade-offs between API cost, latency, and quality; implemented caching and batch inference strategies to maximize GPU utilization.
Multimodal AI GPU Optimization Cost Engineering Node-Based Workflows

Featured Projects

Intel-Graph: Autonomous Multi-Agent Lead Intelligence Engine

Agentic AI • Multi-Agent Systems

Python • LangChain • vLLM • Ollama • Pydantic • Graph Architecture

Problem: Automate lead qualification and intelligence gathering for sales teams using autonomous AI agents.

Solution: Designed a multi-agent system where specialized LLM agents collaborate via a graph-based workflow to research, score, and summarize leads from diverse data sources.

Impact:

  • Reduced manual lead research time by ~70% in pilot testing
  • Implemented dynamic agent routing and state management for complex decision workflows
  • Built evaluation harness to measure agent consistency, factuality, and latency
View Code

Retrieval-Augmented Technical QA System

RAG • NLP • Production ML

Python • BGE-M3 • ChromaDB • FastAPI • Evaluation Frameworks

Problem: Enable accurate, context-aware answers to technical questions from domain-specific documentation.

Solution: Built an end-to-end RAG pipeline with optimized chunking, hybrid retrieval (dense + sparse), and rigorous evaluation using MRR, NDCG, and answer faithfulness metrics.

Impact:

  • Achieved 87.4% Top-5 retrieval accuracy on technical QA benchmark
  • Reduced hallucination rate by 31% via answer grounding and citation enforcement
  • Designed modular architecture for easy domain adaptation and A/B testing
RAG Assistant Finetuning Repo

Technical Skills

Languages & Core Frameworks

Python PyTorch Docker FastAPI Flask SQL (PostgreSQL) MongoDB Hugging Face Transformers React Git Linux

Generative AI & Agentic AI

LangChain LangGraph vLLM RAG OpenAI API CrewAI

Cloud & Infrastructure

AWS (EC2, S3) RunPod Vast.ai

Competitions

Amazon ML Challenge 2025

Top 5% Globally

Python • Ensemble Methods • Feature Engineering • Cross-Validation

Problem: Develop a robust ML solution for Amazon's 2025 challenge dataset with complex feature interactions and imbalanced targets.

Approach: Engineered domain-aware features, implemented stacked ensemble with careful validation strategy, and optimized for generalization under time constraints.

Result: Ranked within top 200 participants globally (~top 5%), demonstrating strong problem decomposition and model selection skills under competition pressure.

View Solution

Side Projects & Explorations

AI Avatar Video Generator

Hobby • Video Synthesis

Python • Diffusion Models • Temporal Consistency

Experimental pipeline for generating personalized AI avatar videos using open-source diffusion models. Focus on maintaining identity consistency across frames.

Demo: Identity-consistent avatar generation
View Code

Automated SOS Messaging via Hand Sign CV

Computer Vision • OpenCV • MediaPipe • Geolocation

Real-time detection of viral TikTok hand signs triggering location-aware emergency messaging—built as a proof-of-concept for accessibility tech.

View Code →

Reddit Sentiment Analyzer

NLP • Transformers • Streamlit

"I don't like to hurt sentiments, so I analyzed them instead." Lightweight tool for aggregate sentiment trend analysis in Reddit communities.

View Code →

Learning Archives

Public notebooks and exercises from my ML/NLP learning journey.

ML/DL Algorithms From-scratch implementations of core algorithms NLP Assignments Tokenization, embeddings, sequence modeling exercises Hugging Face LLM Course Notebooks and quizzes from the specialization

Contact

I'm actively seeking ML Engineering and GenAI roles where I can contribute to building reliable, scalable AI systems. Let's connect.

Email GitHub LinkedIn Hugging Face Resume (PDF)