thejesh-m.github.io

Hi, I'm Thejesh Mallidi.

A
I am a recent Master’s graduate in Data Science from Michigan State University, with a strong focus on building NLP-driven solutions, advanced computer vision systems, and scalable machine learning models.

Skills

Detail-oriented Machine Learning Engineer with expertise in Natural Language Processing, Deep Learning, and Computer Vision. I specialize in architecting end-to-end AI solutions from data pipeline integration to model deployment that deliver measurable impact in real-world applications.

  • Core Expertise: Data Science, Machine Learning, Deep Learning, Natural Language Processing (NLP), Computer Vision, Large Language Models (LLMs), Generative AI & Agentic AI, Predictive Modeling, Decision Analytics, Graph ML
  • Languages: Python, R, C++, C, Bash
  • AI/ML Frameworks & Libraries: PyTorch, TensorFlow, TensorRT, LangGraph, LangChain, LlamaIndex, AutoGen, Phi Data, VectorStores and Retrievers, Hugging Face Transformers, OpenCV, NumPy, Pandas
  • Big Data & Streaming: Hadoop Spark, Apache Kafka
  • Databases: MySQL, PostgreSQL
  • Cloud & MLOps Tools: GCP, AWS, Heroku, Docker, Git, Flask, FastAPI, MLFlow, Nvidia Deepstream

Experience

Quantiphi logo

Quantiphi Analytics

Machine Learning Engineer
  • Engineered an LLM-driven document ingestion pipeline for a leading telecom company processing 2M+ contract documents, extracting unstructured legal text into structured Neo4j graph format while preserving full document hierarchy.
  • Implemented Agentic Contract Assist chatbot (GraphRAG) over the Neo4j graph + document store, improving query relevance by ~45% and reducing retrieval latency from seconds to sub-second responses.
  • Developed and deployed CI/CD workflows on Red Hat OpenShift and GCP Cloud Run, enabling zero-downtime updates and Integrated BigQuery and GCS for scalable contract metadata storage.
  • Created 10+ Re-usable Agentic components those can be used across the organization in various document handling workstreams.
Jun 2025 – Present | Remote, USA
MSU logo

Michigan State University

Graduate Research Assistant
  • Domain-Specific Knowledge Graph: Built a knowledge graph with over 2 million entities and 1.4 million relationships to accelerate hypothesis generation in plant science, leveraging NLP techniques such as Named Entity Recognition and Relation Extraction.
  • LLM-Driven Entity Extraction from Scientific Texts: Achieved 94% entity extraction accuracy by Finetuning state-of-the-art domain specific BERT models for entity extraction, enabling robust identification of relevant biological entities.
  • Custom BERT-Based Relation Extraction: Designed and implemented a custom BERT-based multi-relation extraction model to identify and extract multiple relationships within a single sentence.
  • LLM-Powered Knowledge Graph Querying: Enabled accurate question answering over a knowledge graph using GenAI Agents and Retrieval-Augmented Generation (RAG) techniques, optimized through prompt engineering.
  • Domain-Adaptive BERT Pretraining: Pretrained a BERT model on 400K+ plant science PubMed abstracts with selective masking strategy and weighted-loss to prioritize ~30K key plant terms. Achieved significantly lower MLM validation loss compared to baseline BERT.
Jan 2024 – Present | East Lansing, MI, USA
Quantiphi logo

Quantiphi Analytics

Machine Learning Engineer
  • Responsible for designing and implementing Automatic Information Extraction systems by combining custom Computer Vision and NLP LLMs in the BFSI domain.
  • Delivered solutions for 10+ Proof of Concepts (PoCs)b> on document classification & extraction, object detection, image classification, and semantic segmentation tasks.
  • Built end-to-end Airflow pipelines to orchestrate object detection, classification, and segmentation tasks, supporting scalable and reproducible ML model training.
  • Developed insurance-focused predictive models for risk assessment and claim outcome prediction Logistic Regression, Decision Trees, Bagging, Random Forest, and Gradient Boosting Machines (GBM).
  • Developed and implemented algorithms using Machine Learning and Statistical Modeling techniques to improve performance, quality, data management, and accuracy.
  • Built and optimized real-time inference pipelines for Nvidia edge devices using DeepStreamb> with custom C++ plugins, achieving 27 FPSb> for latency-sensitive applications.
  • Deployed and monitored 20+ Machine Learning models in production using AWS SageMaker, GCP Cloud Run & Kubernetes, Flask,and FastAPI.
Jun 2021 – Jul 2023 | Bangalore, India

Projects

Screenshot of Pre-training GPT model project
Pre-training GPT Model

This project provides end‑to‑end code for training a GPT‑style language model from scratch (or fine‑tuning an existing checkpoint) on large text corpora. It includes utilities for downloading and preprocessing data, a configurable training pipeline built on PyTorch .

Accomplishments
  • Re-implemented the full decoder-only Transformer in ~300 lines of model.py and a matching ~300 line training loop—embracing the “small, clean, interpretable” style while adding modern upgrades such as FlashAttention, gradient-checkpointing, and AMP/bfloat16 support for memory-efficient scaling.
  • Scaled training for multi-GPU: Orchestrated PyTorch DDP for distributed training across nodes, streaming the 38 GB OpenWebText corpus through a custom shard-loader to pre-train a 124 M-parameter GPT model.
  • Implemented configurable Top-k/Top-p sampling in the generation CLI: Giving users fine-grained control over randomness and diversity when producing sample continuations.
Screenshot of MyHealthAgent analysis project
MyHealthAgent: Personalized Multi‐Agent Nutrition Assistant

A GenAI‐powered, multi‐agent nutrition assistant designed to help users log meals, extract medical conditions from lab‐report images, and receive real‐time, personalized dietary guidance.

Accomplishments
  • Engineered a GenAI‐powered nutrition assistant that analyzes user chronic conditions and meal logs to deliver real‐time, personalized dietary recommendations in under 2 seconds.
  • Formulated condition-to-nutrition pipeline that extracts user medical conditions from lab report images through OCR and a LLaMA-based model, then computes 10+ macro- and micro-nutrient thresholds tailored to those conditions.
  • Designed a Web App that allowed users to chat with a Multi-Agent System for condition-specific food recommendations, dietary planning, and food image classification with ~90 % ingredient-recognition accuracy.
Screenshot of road surface analysis project
SurfaceVision: Road Damage Analytics

Developed a comprehensive methodology to identify deflected or damaged regions in road surfaces, enhancing road maintenance strategies with volumetric quantification of surface deformities.

Accomplishments
  • Utilized advanced data analysis, image analysis, and visualization techniques to assess and quantify road damage.
  • Applied cutting-edge technologies, including Python, Vedo, and deep learning models like PointNet++ for accurate surface modeling.
  • Enhanced road maintenance strategies by enabling volumetric quantification of surface deformities.
Screenshot of Image Captioning using Attention Models
Image Captioning Using Attention Models

This project aims to describe the activity in a given image. The dataset used is the Flickr 8k dataset along with pre-trained models like VGG16 and ResNet for image representation.

Accomplishments
  • Utilized Natural Language Processing (NLP) techniques combined with deep learning models (CNN & LSTM).
  • Implemented attention mechanisms to improve the quality of captions generated for images.
  • Evaluated the model using metrics like BLEU and achieved competitive results.
  • Tech Stack: NLP, Deep learning (CNN & LSTM), keras, Python
Screenshot of Semantic Segmentation for Self-Driving Cars
Semantic Segmentation for Self-Driving Cars

Self-driving cars require a deep understanding of their surroundings. This project focuses on semantic segmentation using neural networks to recognize roads, pedestrians, cars, and sidewalks with pixel-level accuracy.

Accomplishments
  • Developed a neural network using UNet and SegNets for semantic segmentation tasks.
  • Optimized the model to assign each pixel to a target class such as road, car, pedestrian, or sign.
  • Leveraged TensorFlow and Keras frameworks for implementation and evaluation.
Screenshot of SaveSphere Project
SaveSphere: Anamoly Action Detection

SaveSphere is a real-time video analysis system that identifies potential threats and sends email alerts for prompt action.

What it does
  • Identifying Potential Threats: SaveSphere actively recognizes possible dangers or risky situations through real-time video analysis using object detection techniques.
  • Email Alert Notifications: Once a threat is detected, SaveSphere swiftly sends email alerts to facilitate prompt action.
How we built it

To ensure the success of the project, two fundamental components needed implementation:

  • The initial phase involved training the object detection model to identify threats in videos using Python and PyTorch.
  • The subsequent phase focused on integrating the model into a web platform and enabling real-time email notifications for user safety.

For seamless integration into a web platform, Flask, React, and CSS were utilized to collectively bring the project to fruition.

Education

Michigan State University

East Lansing, Michigan

Degree: Masters in Data Science
Duration: 08/2023 - 04/2025
CGPA: 3.95/4.0

Relevant Coursework:

  • Advanced Machine Learning
  • Statistical Modeling & Data Analysis
  • Computer Vision
  • Natural Language Processing
  • Computional Optimization

Madanapalle Institute of Technology and Science

Madanapalle, India

Degree: BTech - Computer Science & Engineering
Duration: 07/2017 – 05/2021
CGPA: 9.23/10

Relevant Coursework:

  • Data Structures and Algorithms
  • Database Management Systems
  • Operating Systems
  • Python, C & Java

Contact