Skip to content
Manish Maurya

Open to work · Immediate joiner

Manish Maurya

AI/ML Engineer, Generative AI & LLMs

I build Generative AI systems with large language models: fine-tuning with LoRA, QLoRA and quantization, RAG-based data pipelines, prompt engineering and AI agent evaluation. I'm currently a Research Intern at LTRC, IIIT Hyderabad.

View projectsDownload résumé

About

What I bring to an AI team

I'm an AI/ML engineer focused on Generative AI and large language models. I fine-tune open LLMs with parameter-efficient methods (LoRA, and QLoRA with 4-bit quantization) in Hugging Face Transformers, build RAG-based pipelines that generate training and evaluation data, and write prompts that get consistent, checkable output from LLMs.

I care about whether things actually work, so I measure them. One of my open-source projects benchmarks security tools for AI agents against what really executed; another validates every sentence an LLM generates and runs 37 automated tests in CI.

I'm a Research Intern at LTRC, IIIT Hyderabad, and before that I interned at DRDO's Centre for Artificial Intelligence and Robotics, where I benchmarked NER models and worked on a model-agnostic inference engine with ONNX Runtime. I have a B.Tech in Computer Science and Information Technology from REVA Institute of Technology (2025).

  • LLM fine-tuning

    Adapt open LLMs to new tasks and languages with LoRA and QLoRA, using 4-bit quantization to train within limited GPU memory.

    • LoRA
    • QLoRA
    • Quantization
    • SFT
    • Hugging Face

    Proof: LTRC internship

  • RAG & synthetic data

    Build RAG-based pipelines that pair LLM generation with heuristic filters to produce clean fine-tuning and evaluation datasets.

    • RAG
    • Synthetic data
    • Ollama
    • Data quality

    Proof: Synthetic Data Generator, LTRC internship

  • Prompt engineering & evaluation

    Design prompts for local and API-served models, then check the output automatically: validation rules, deduplication and side-by-side model comparisons.

    • Prompt engineering
    • LLM evaluation
    • Model comparison

    Proof: Synthetic Data Generator

  • AI agents & MCP

    Understand how tool-using agents fail (tool poisoning, rug pulls, data leaks) and test defences against ground truth.

    • AI agents
    • MCP
    • LLM security

    Proof: MCP-GuardBench

  • Packaging & deployment

    Turn models and tools into APIs, CLIs and containers, with tests and CI, and publish them as installable packages.

    • FastAPI
    • Docker
    • ONNX Runtime
    • GitHub Actions
    • PyPI

    Proof: MCP-GuardBench, DRDO CAIR

  • Document AI data

    Generate labelled document-layout datasets (PDFs, page images, COCO and YOLO boxes) with a local LLM, for training layout models.

    • Document AI
    • Ollama
    • ReportLab

    Proof: Synthetic Document Factory

Projects

Selected work

Open-source projects with the source on GitHub. Each case study covers the problem, the architecture, the engineering decisions and the limitations, and every number quoted comes from the repository.

  • AI agents · LLM security · Evaluation

    MCP-GuardBench

    A reproducible lab and CLI for Model Context Protocol security. It checks the MCP servers you use for tool poisoning and rug pulls, and benchmarks security tools against what actually executed.

    • 7/9attacks detected
    • 8/9attacks blocked
    • 0/3false positives
    • Python
    • FastAPI
    • Pydantic
    • SQLAlchemy
    • PostgreSQL
    • Docker

  • Generative AI · Synthetic data · Multilingual NLP

    Multilingual Synthetic Data Generator

    A CLI that builds clean, deduplicated sentence datasets in 13 South Asian and global languages with any local Ollama model, plus a document generator that writes structured English or Hindi articles as JSON and PDF.

    • 13languages
    • 37pytest tests in CI
    • Python
    • Ollama
    • REST APIs
    • ReportLab
    • pytest
    • GitHub Actions

  • Generative AI · Document AI · Synthetic data

    Synthetic Document Factory

    Generates labelled document-layout datasets with a local LLM: PDFs, page images and exact bounding boxes in COCO and YOLO format, using DocLayNet's label set.

    • COCO + YOLOboxes recorded while rendering
    • 11DocLayNet label ids
    • Python
    • Ollama
    • ReportLab
    • Matplotlib
    • pypdfium2
    • pytest

  • Computer vision · Detection and tracking

    Basketball Video Analytics

    A computer vision pipeline that detects and tracks players and the ball in game footage, assigns teams, estimates possession, passes and interceptions, and projects players onto a top-down tactical court.

    • Python
    • Ultralytics YOLO
    • ByteTrack
    • Fashion-CLIP
    • OpenCV
    • Docker

Research

Research and investigations

Internship research and independent work, each labelled with its status.

  • Ongoing research

    Fine-tuning LLMs and zero-shot NER for low-resource Indic languages

    LTRC, IIIT Hyderabad · May 2026 – present

    Research internship work on adapting transformer LLMs to Indic languages that have little training data, and on the data needed to fine-tune and evaluate them.

    • LoRA/QLoRA fine-tuning workflows for transformer LLMs in Hugging Face Transformers.
    • RAG-based synthetic data pipelines that pair LLM generation with custom heuristics to build fine-tuning and evaluation datasets.
    • Zero-shot named entity recognition with GLiNER.
    • LoRA
    • QLoRA
    • RAG
    • GLiNER
    • Indic NLP

  • Completed · unpublished

    Comparative study of monolingual and multilingual NER

    DRDO CAIR, Bengaluru · Feb – Apr 2025

    Internship study comparing NER architectures, alongside work on a model-agnostic NER inference engine.

    • Benchmarked monolingual and multilingual NER architectures, with F1-score as the primary metric.
    • Investigated zero-shot, entity-agnostic NER with span-based frameworks such as GLiNER.
    • Engineered language-ID routing, standardized output schemas and multi-process ONNX inference for the inference engine.
    • NER
    • GLiNER
    • F1 evaluation
    • ONNX Runtime

  • Independent research

    Evaluating MCP security controls against ground truth

    Open source

    A benchmark that scores LLM-agent security tools on whether attacks were actually stopped, read from each test server's own record of what executed, rather than on the tools' own reports.

    • LLM agents
    • Security
    • Evaluation methodology
    Case study Repository

  • Learning notes

    Data science foundations

    Handwritten notes, 130+ pages

    Notes on statistics and probability, ML foundations, dimensionality reduction (PCA) and NLP text representations (bag of words, TF-IDF, Word2Vec), written to understand why methods work, not just how to call them.

    Read the notes

Skills

Technical skills

Grouped by discipline, with links to where each group was used.

  • Generative AI & LLMs

    • LLM fine-tuning
    • LoRA
    • QLoRA
    • Quantization (4-bit)
    • Supervised fine-tuning
    • RAG
    • Prompt engineering
    • LLM evaluation
    • Synthetic data generation
    • AI agents & MCP
    • Hugging Face Transformers
    • Ollama

    Used in LTRC internship, MCP-GuardBench, Synthetic Data Generator

  • Backend, deployment & tools

    • FastAPI
    • Pydantic
    • SQLAlchemy
    • PostgreSQL
    • Docker
    • ONNX Runtime
    • REST APIs
    • Git & GitHub
    • GitHub Actions
    • pytest

    Used in MCP-GuardBench

  • Machine learning & deep learning

    • PyTorch
    • TensorFlow
    • scikit-learn
    • Pandas
    • Data preprocessing
    • Model evaluation

    Used in DRDO CAIR NER benchmarking

  • NLP

    • Named entity recognition
    • Zero-shot NER (GLiNER)
    • Multilingual & Indic NLP

    Used in LTRC and DRDO CAIR internships

  • Programming languages

    • Python
    • Java

  • Computer vision

    • OpenCV
    • Ultralytics YOLO
    • ByteTrack
    • Roboflow

Experience

Experience and education

  1. Research Intern · LTRC, IIIT Hyderabad

    May 2026 – Present

    Language Technologies Research Centre · Internship

    • Develop LoRA/QLoRA fine-tuning workflows for transformer LLMs in Hugging Face Transformers, targeting low-resource Indic languages.
    • Design RAG-based synthetic data pipelines that pair LLM generation with custom heuristics to build fine-tuning and evaluation datasets.
    • Advance zero-shot named entity recognition (NER) research using GLiNER.
    • Automate data collection and experiment workflows by integrating third-party and model-serving APIs.
    • LoRA
    • QLoRA
    • Hugging Face Transformers
    • RAG
    • GLiNER

  2. Research Intern · DRDO CAIR

    Feb 2025 – Apr 2025

    Centre for Artificial Intelligence and Robotics, Ministry of Defence, Govt. of India · Internship · Bengaluru

    • Benchmarked monolingual and multilingual NER architectures in a comparative study, using F1-score as the primary evaluation metric.
    • Investigated zero-shot, entity-agnostic NER paradigms using span-based frameworks such as GLiNER.
    • Engineered language-ID-based routing, standardized output schemas, and ONNX-based multi-process inference for a scalable, model-agnostic NER inference engine.
    • NER
    • GLiNER
    • ONNX Runtime
    • Multilingual NLP

Education

B.Tech, Computer Science and Information Technology

REVA Institute of Technology, Bangalore

2021 – 2025

Relevant coursework: Machine Learning, NLP, Data Structures & Algorithms, DBMS, Operating Systems, Computer Networks

Certifications

DeepLearning.AI
Machine Learning Specialization
PyTorch for Deep Learning Professional Certificate
Transformers in Practice
Agentic AI
TensorFlow Developer Professional Certificate
IBM
Python for Data Science
Johns Hopkins University (Coursera)
HTML and CSS

Contact

Get in touch

I'm looking for full-time AI/ML, Generative AI and LLM engineering roles in Bengaluru, Hyderabad or remote, and can join immediately. Email is the quickest way to reach me.

manish.maurya0408@gmail.com