Omkar
Thorve.

I build AI systems from research to deployment. Some become papers. Some go live. A few do both.

About

The engineer who crossed disciplines.

Mechanical engineer turned AI/ML researcher. I train models, build pipelines, and ship products. Not the “I use the ChatGPT API” kind.

I care about the gap between a model that works on Colab and one that holds up on real sensor data. That's where most of my work lives.

⚙️Ex-mechanical engineer🏀Zonal basketball captain📝Writes on Hashnode🔬Currently: Production RAG🏭AI × industrial systems

Impact by the numbers

2–4 hrs
Diagnosis time cut to 77 s
3
Systems in production
20%
Inference cost reduced
2
Peer-reviewed papers

Currently exploring

Reinforcement learning, machine vision, and AI-driven industrial automation. Also building production-grade generative AI apps, not wrappers. Full systems.

“Mechanical engineering taught me how systems fail. AI taught me how to predict it before they do.”
Experience

Where I've built things.

John Deere India Pvt. Ltd.

AI/ML Intern

Jun 2025 – Dec 2025
Pune · On-site
  • Applied object detection & segmentation models (YOLO, SAM2, DETR, ViT, Grounding DINO) to automate manual processessaved 4+ hrs/task
  • Fine-tuned Large Language Models on custom image datasets50% accuracy improvement, hallucinations reduced
  • Built end-to-end speech-to-text pipeline: Whisper + GPT-4o Transcribe + GPT-4o Mini with JS frontend/backend10% transcription boost, saved 2 hrs/day
  • Designed multi-agent workflow using Knowledge Graphs + GNN, deployed on AWS (EC2, S3, DynamoDB)4+ hours saved in document workflows
  • Implemented cross-encoder reranker for semantic retrieval outperforming similarity search. Databricks for large-scale extraction + OpenAI embeddings for evaluation
  • Conducted hallucination-reduction experiments for deployed LLMsimproved reliability & robustness
PythonYOLOSAM2LangGraphAWSGPT-4oDatabricksGNNLangChain

K.S.J Recruitment Pvt. Ltd.

Data Analyst Intern

Feb 2023 – Feb 2024
Remote
  • Created and maintained Power BI dashboards for recruitment metrics, providing actionable insights to optimize hiring strategies
  • Performed data cleaning and analysis using Excel and Python ensuring accurate reporting
  • Automated data workflows20% reduction in manual effort
  • Collaborated with cross-functional teams to streamline recruitment data pipelines
PythonPower BIExcelSQL
Projects

Here's what I've been building.

https://rca-frontend-dri9.onrender.com/

Flagship project

Agentic Industrial Diagnostics

What if a maintenance engineer could sleep while four AI agents diagnosed the breakdown?

Under Review, Springer

Multi-agent LLM system for real-time root cause analysis of industrial equipment failures. An LSTM Autoencoder catches anomalies with 95.3% precision, an OWL/RDF knowledge graph maps causal chains, and four LangGraph agents (Diagnostic, Reasoning, Planning, Learning) collaborate to produce an explanation and remediation plan in ~77 seconds.

FastAPI · LangGraph · LangChain · Gemini · LSTM · OWL/RDF · Next.js 14

Outcomes

  • 84.6% root cause identification rate on AI4I dataset
  • 77 s avg. diagnosis, down from 2-4 h manual
  • 72% zero-shot cross-domain transfer (AI4I to MetroPT)
  • 100% workflow success rate

More work

  • 01

    SwishFit

    LiveNov 2025

    AI-powered basketball training platform where Google Gemini 2.5 Flash generates personalised workout plans. Role-based access for players, coaches, and admins. JWT + bcrypt + Helmet.js security stack. Deployed on Vercel + Render + MongoDB Atlas.

    React (Vite) · Node.js · Express · MongoDB Atlas · Gemini 2.5 · JWT

  • 02

    GangaFlow

    IEEE ICoICC 2025Dec 2024

    AI-driven water quality monitoring for the Ganga River Basin. YOLOv8 bounding box detection, U-Net area segmentation, and AlexNet severity classification fused into a single real-time pipeline giving environmental agencies actionable, annotated outputs.

    Python · YOLOv8 · U-Net · AlexNet · OpenCV

  • 03

    CNC Anomaly Detector

    Jan 2025

    Convolutional autoencoder trained on tri-axial vibration signals from CNC milling machines. Dynamic threshold adapts to machine state drift without retraining. Ships with a live Streamlit dashboard: upload a CSV, get a verdict.

    Python · Autoencoder NN · Streamlit · Signal Processing

  • 04

    Mild Steel Degradation

    1st Runner-up, Ninja Hack 2K252025

    Real-time corrosion detection on microscopic steel imagery using CNN + HSV colour segmentation + Canny edge detection. Flask web app with corrosion zone heatmap overlays and automated severity report generation.

    Python · CNN · OpenCV · Flask

  • 05

    Hotel Reservation MLOps

    Apr 2025

    End-to-end MLOps workflow: MLflow tracks every experiment, FastAPI serves predictions behind API key auth. Ships in Docker Compose.

    Python · MLflow · FastAPI · Random Forest · Docker · scikit-learn

Publications

Research contributions.

GangaFlow: A Multi-Model Deep Learning Framework for Real-Time River Pollution Detection and Analysis Using Drone Imagery

IEEE ICoICC 2025View paper

Agentic Industrial Diagnostics: A Multi-Agent LLM Framework for Explainable Root Cause Analysis

Under Review, Springer
Achievements

Outside the notebooks.

🏆

1st Runner-up

The Great Ninja Hack 2K25

DYPCET Kolhapur

Mild Steel Degradation Analysis Using Microscopic Imaging & Deep Learning

🥇

1st Position

Smart India Hackathon

Internal, SIT

Traffic Signal Control System using YOLOv5 + Reinforcement Learning

🥈

2nd Runner-up

Best Manager Competition

NICMAR University

Competed in the prestigious national-level management challenge

🏀

Zonal Champion

Basketball, Team Captain

Deogiri Institute

Led college team to zonal championship. Competed at State & National level.

Not everything I build is code.

Recommendations

What people say.

From colleagues, managers, and mentors, in their own words.

I guided Omkar through his M.Tech dissertation project over the past year, with a focus on agentic AI for predictive maintenance closely aligned with my own research area. What stood out was his ability to work independently. He took ownership of both the research and implementation, showing strong research instincts and a genuinely hardworking attitude. Our collaboration resulted in a paper currently under review with Springer. He has deep knowledge of agentic frameworks and systems, and the kind of intellectual curiosity that drives good research. I'd genuinely recommend him for roles in AI/ML, he's the kind of student you don't have to push.

PK

Dr. Pooja Kamat

Associate Professor (AI and ML) | Research Associate, SCAAI · Symbiosis Institute of Technology

Pooja was Omkar's mentor

I had the chance to work closely with Omkar during our M.Tech, where we collaborated on projects, patents, and hackathons together. He's a dedicated and skilled person who always brings good ideas and a calm problem-solving approach to the team. It was great continuing the journey as colleagues at John Deere as well. I'd definitely recommend him to anyone looking for a reliable and talented professional.

NB

Nishit Bohra M

AI Engineer · LLMs · RAG · Agentic AI | IEEE Author · Patent Holder · John Deere

Colleague at M.Tech & John Deere

I had the privilege of collaborating with Omkar during his internship, where he contributed to several AI and automation projects. His energy and eagerness to learn were remarkable, especially in exploring and implementing GenAI technologies and making them practically workable. As an intern, grasping complex concepts and applying them effectively is never easy, yet Omkar managed it with confidence. Though he started with limited business knowledge, he quickly picked up the essentials and worked towards delivering solutions aligned with business needs. Omkar's adaptability, curiosity, and commitment to learning made him a valuable contributor, and I am confident he will continue to excel in his career.

LD

Latheef D

Data Scientist · John Deere

Project lead on the 2D drawing automation tool

Full recommendations on LinkedIn

Education

Where the foundation was built.

M.Tech in Artificial Intelligence and Machine Learning

Symbiosis Institute of Technology

2024 – 2026

In Progress

Coursework

Supervised & Unsupervised MLDeep LearningMachine VisionNLPReinforcement LearningGraph Neural NetworksGANsLLMsAgentic AI

B.Tech in Mechanical Engineering

Deogiri Institute of Engineering and Management Studies, Aurangabad

2020 – 2023

Completed

Coursework

Image ProcessingLinear AlgebraCalculusNumerical AnalysisStatisticsControl SystemsPythonLinux

Certifications

Generative AI Engineering with LLMs

IBM

Data Analysis and Visualization with Power BI

Microsoft

AWS Cloud Training

Unnati Development Pvt Ltd

Writing

Things worth writing down.

I write about what I'm learning: RAG pipelines, ML systems, things that took me too long to figure out.

RAG · LLMs

Getting Started with RAG: Building a Document Q&A System

How to build a retrieval-augmented generation pipeline from scratch: chunking, embeddings, and why similarity search isn't always enough.

Agentic AI · Real-world

What I Learned Building a Multi-Agent System at John Deere

The gap between a multi-agent demo and a production-grade system is larger than the papers make it seem. Here's what actually surprised me.

Career · Learning

From Mechanical Engineering to ML: The Honest Version

Not a LinkedIn pivot story. A practical account of what transferred, what didn't, and what I had to relearn from scratch.

Read all articles on Hashnode
Contact

Let's build something.

I'm open to ML engineering roles, research collaborations, and genuinely interesting problems. Drop me a line. I read every message.

Download Resume