Madhumitha Murthy
Available for full-time roles from July 2026 Ā· Singapore

Madhumitha
Murthy

AI Engineer Ā· Agentic AI Ā· LLMs Ā· NLP Ā· Production ML

I build end-to-end AI systems ā€” agentic pipelines, RAG, fine-tuned LLMs, anomaly detection, full-stack fintech with LLM explainability, and 5G sensing simulations. Production-ready with FastAPI, Docker, Kubernetes, and GitHub Actions CI/CD. 5 peer-reviewed publications. MSc at NTU Singapore, graduating June 2026.

5 Publications
0.84 Macro-F1 Ā· EACL
3rd Global Ā· NLP Task
70K Competitors Beat Ā· Zoho
Get in touch View Projects ā†’ Download CV

What I've Built

Projects & Systems

Full-Stack AI
Fintech Payment Risk Dashboard
Full-stack fintech application with 4 Spring Boot microservices (User, Transaction, Risk, Intelligence) and a React dashboard; services communicate via REST, persisted to MongoDB, fully containerised with Docker Compose. Hybrid ML Risk Service bridges Java to a Python LSTM Autoencoder anomaly detection API: weighted scoring (rule-based 40% + ML 60%), auto-flags HIGH-risk transactions. GenAI Intelligence Service uses Groq API (Llama 3.1) for plain-English transaction explainability and NL ā†’ structured filter query parsing.
Java Spring Boot React MongoDB Python FastAPI LSTM Autoencoder Groq LLM Docker GitHub Actions
ā†’ 4 microservices Ā· hybrid ML+rule risk scoring Ā· LLM explainability Ā· CI/CD multi-stage Docker build Ā· GitHub ā†’
Agentic AI
Agentic AI Tech News Pipeline
End-to-end agentic pipeline that autonomously fetches articles from ArXiv, HackerNews, TechCrunch, VentureBeat, and AI research blogs; filters by relevance; summarises with Gemini 2.0 Flash; and delivers a styled HTML digest to Gmail at 11:55 PM SGT daily. Built-in RAG query interface using ChromaDB + Sentence Transformers ā€” ask questions over stored articles in natural language.
Agentic AI Gemini 2.0 Flash ChromaDB RAG Docker Kubernetes GitHub Actions
ā†’ 97 articles fetched daily Ā· semantic RAG query interface Ā· CI/CD + K8s CronJob deployment Ā· GitHub ā†’
Production ML
RAG-Powered Document Q&A API
Upload any PDF ā†’ section-aware chunking with NL context prefixes ā†’ HuggingFace sentence-transformer embeddings ā†’ FAISS semantic search ā†’ Groq Llama 3.3-70B grounded answers via FastAPI. 3 retrieval configs benchmarked in MLflow; PDFs and index persisted to AWS S3. Containerised with Docker.
LangChain FAISS HuggingFace FastAPI Docker MLflow AWS S3
ā†’ 64% keyword recall Ā· 0.63s latency Ā· section-aware chunking + top-k tuning improved recall from 28% to 64% Ā· GitHub ā†’
Production ML
Domain-Specific LLM Fine-Tuning with LoRA
Fine-tuned DistilBERT on FinancialPhraseBank for 3-class sentiment using LoRA (PEFT) ā€” 98.7% fewer trainable parameters (887K vs 67M full fine-tune). 3 MLflow experiments tracked across learning rate and LoRA rank. Deployed via FastAPI REST API on AWS EC2; adapter checkpoints stored in AWS S3, containerised with Docker.
LoRA / PEFT DistilBERT HuggingFace FastAPI Docker MLflow AWS (EC2, S3)
ā†’ F1 = 0.846 Ā· Accuracy = 89.2% Ā· 98.7% parameter reduction vs full fine-tune Ā· GitHub ā†’
Production ML
Time-Series Anomaly Detection Pipeline
Anomaly detection on NASA SMAP satellite sensor data (54 channels, 562 anomaly sequences) ā€” LSTM Autoencoder achieved F1 = 0.737, AUC-ROC = 0.857, FAR = 0.7%. Anomaly segment inspection formulated as a fractional knapsack linear programme (SciPy linprog / HiGHS); LP captured +17.6% more anomaly signal than naive greedy on a 10% budget. SimPy discrete event simulation of downstream inspection workflow: LP-prioritised vs greedy schedules ā€” LP reduced makespan by 24% and improved machine utilisation from 50% to 76%. KS-test drift monitor flags distribution shift at p < 0.05; /drift/status endpoint for retraining triggers. 150-test suite with GitHub Actions CI/CD.
LSTM Autoencoder PyTorch SciPy linprog SimPy scikit-learn FastAPI Docker MLflow
ā†’ F1 = 0.737 Ā· AUC-ROC = 0.857 Ā· LP +17.6% signal vs greedy Ā· SimPy 24% makespan reduction Ā· ~12ms inference Ā· GitHub ā†’
Industrial Defect Detection Pipeline
End-to-end defect detection and characterisation pipeline on the MVTec Anomaly Detection dataset (crack, cut, hole, print defect types). scikit-image preprocessing: Gaussian denoising, Otsu thresholding, morphological cleaning, regionprops measurement. OpenCV for contour detection and mask post-processing. Fine-tuned PyTorch ResNet18 for binary classification, tracked in MLflow. Deployed as FastAPI REST API.
PyTorch ResNet18 scikit-image OpenCV MLflow FastAPI
ā†’ >90% classification accuracy Ā· regionprops defect characterisation Ā· FastAPI inference API Ā· GitHub ā†’
Instance Segmentation ā€” Mask R-CNN
Implemented Mask R-CNN from pretrained COCO weights for simultaneous bounding-box detection and pixel-level instance segmentation. OpenCV preprocessing: resizing, normalisation, morphological filtering. Extended with scikit-image post-processing: Otsu thresholding, regionprops-based region measurement (area, perimeter, eccentricity, solidity) on each detected mask.
TensorFlow Mask R-CNN OpenCV scikit-image NumPy
ā†’ Pixel-level instance segmentation Ā· regionprops defect characterisation Ā· COCO pretrained Ā· GitHub ā†’
5G ISAC Digital Twin Simulation
Built a physics-accurate 3D urban digital twin (Fusionopolis, Singapore) using NVIDIA Sionna ray tracing. Implemented OFDM radar with Static Clutter Cancellation at 59 GHz, Doppler-based trajectory prediction at 0.333s intervals, and sensing-aided beamforming.
NVIDIA Sionna Python NumPy/SciPy Blender 5G NR
ā†’ First-authored IEEE ISCAS 2026 Ā· Co-authored Nature npj Wireless Tech Ā· GitHub ā†’
Fake News Detection in Dravidian Languages
Fine-tuned XLM-RoBERTa, mBERT, and ALBERT for multilingual misinformation detection in Dravidian languages. Achieved Macro-F1 of 0.84 and 3rd place on the DravidianLangTech@EACL 2024 shared task leaderboard. Published at DravidianLangTech Workshop, EACL 2024.
XLM-RoBERTa mBERT ALBERT HuggingFace Python
ā†’ Macro-F1: 0.84 Ā· 3rd place globally Ā· EACL 2024 Ā· GitHub ā†’
Object Detection Lab (Mask R-CNN)
Developed full lab manual and reference implementation for object detection and instance segmentation using Mask R-CNN for NTU postgraduate students.
Mask R-CNN Computer Vision PyTorch
ā†’ Adopted in NTU postgrad curriculum Ā· GitHub ā†’

Technical Stack

Skills

Agentic AI
Agentic Pipelines AI Agents Generative AI RAG ChromaDB FAISS Vector Database Sentence Transformers RAGAS Vector Embeddings Retrieval Evaluation Prompt Engineering
ML / Deep Learning
PyTorch TensorFlow HuggingFace Transformers scikit-learn LLMs LSTM Transformer
NLP
Transformer Fine-Tuning LoRA / PEFT Multilingual NLP LLM Guardrails XLM-RoBERTa mBERT
MLOps / Deploy
FastAPI REST API Docker Kubernetes GitHub Actions CI/CD MLflow Model Versioning Drift Detection Model Monitoring Retraining Automation Experiment Tracking APScheduler LangChain A/B Testing Git HuggingFace Spaces AWS (EC2, S3)
Scientific Computing
NumPy Pandas SciPy Matplotlib
Optimisation & Simulation
Linear Programming (LP) Fractional Knapsack Constrained Optimisation SciPy linprog (HiGHS) Genetic Algorithms Discrete Event Simulation (SimPy) MTTF/MTTR Modelling
Programming
Python MATLAB SQL Kotlin
Signal Processing
OFDM Radar 5G/6G NR ISAC Beamforming Delay-Doppler Processing Channel Modelling
Simulation
NVIDIA Sionna Ray Tracing Blender OpenStreetMap Digital Twin

Where I've Worked

Experience

Research Attachment
Oct 2024 ā€“ Dec 2025
Institute for Infocomm Research (IĀ²R), A*STAR Ā· Singapore
  • Built a high-fidelity 3D ISAC simulation platform using NVIDIA Sionna ray tracing, modelling multipath propagation, Doppler effects, and dynamic occlusions for vehicle sensing in a realistic urban environment (Fusionopolis, Singapore)
  • Constructed a 3D digital twin by integrating OpenStreetMap road geometry and Blender-modelled buildings, enabling physics-consistent signal modelling across monostatic and bistatic radar configurations
  • Designed OFDM radar algorithm with Static Clutter Cancellation (SCC) at 59 GHz / 100 MHz bandwidth, suppressing static reflections to isolate moving vehicle targets with distinct Doppler signatures
  • Developed Doppler-based trajectory prediction algorithm achieving close vehicle tracking at update intervals as short as 0.333s, enabling proactive beam alignment in high-mobility scenarios
  • Implemented sensing-aided beamforming using predicted vehicle positions to update beamforming weights in real time, achieving BER comparable to perfect-DOA benchmarks at lower computational cost
  • First-authored IEEE ISCAS 2026; co-authored npj Wireless Technology (Nature Portfolio)
Research Assistant
Sept 2025 ā€“ Dec 2025
NTU College of Computing & Data Science Ā· Singapore
  • Developed specifications for a bias-resilient peer evaluation system for NTU
  • Conducted literature review on rubric-based peer evaluation methodologies
Postgraduate Teaching Assistant
Aug 2024 ā€“ Jan 2025
NTU School of Electrical & Electronic Engineering Ā· Singapore
  • Developed lab manual and reference implementation for object detection using Mask R-CNN
  • Graded coursework for IE2108: Data Structures & Algorithms in Python (2 groups)
Android Project Trainee
Feb 2024 ā€“ Mar 2024
Zoho Corporation Ā· India
  • Built Kotlin-based banking application with account management and transaction features using Coroutines and asynchronous workflows
  • Won 1st place at Cliq-Trix 2024 out of 70,000 participants (bot-building competition); awarded ā‚¹1,00,000 cash prize and offered a full-time position at Zoho

Peer-Reviewed Research

Publications

01
IEEE ISCAS Ā· 2026
ISAC for Intelligent Transportation: Ray-Tracing, Clutter Cancellation, and Sensing-Aided Beamforming
5G-based Integrated Sensing and Communication for intelligent vehicle systems using ray-tracing simulation.
Accepted
02
npj Wireless Technology Ā· Nature Portfolio
Integrated Communication and Sensing: Algorithms & 3D Simulation Insights
End-to-end 3D digital twin simulation for ISAC systems with algorithmic performance analysis.
Accepted
03
DravidianLangTech Workshop Ā· EACL 2024
Fake News Detection Using Deep Learning Models in Dravidian Languages
Fine-tuned XLM-RoBERTa, mBERT, and ALBERT for multilingual misinformation detection. Macro-F1: 0.84 Ā· 3rd place globally.
Published
04
CEUR Workshop Proceedings Ā· FIRE 2023
Sarcasm Detection in Dravidian Languages using Transformer Models
Deep learning pipeline for sarcasm identification in code-mixed social media text. Macro-F1: 0.70 / 0.68.
Published
05
LT-EDI Workshop Ā· ACL 2023
Transformer Models to Detect Levels of Depression from Social Media Text
Fine-tuned ALBERT and RoBERTa for mental health classification on social media data.
Published

Academic Background

Education

MSc in Signal Processing & Machine Learning
Nanyang Technological University, Singapore
Aug 2024 ā€“ Jun 2026 (Expected)
School of Electrical & Electronic Engineering. Dissertation with IĀ²R, A*STAR on ISAC using 5G for V2V/Infrastructure Communication and Localisation.

Coursework: Genetic Algorithms & ML, NLP, AI & Data Mining, Pattern Recognition & Deep Learning, Real-Time DSP, Machine Vision.
BE in Computer Science & Engineering
Meenakshi Sundararajan Engineering College, Anna University
Sept 2020 ā€“ July 2024
Chennai, India. First-class honours.

Coursework: Data Structures & Algorithms, Software Engineering, Database Systems, Computer Networks.

Recognition

Achievements

1st Place ā€” Cliq-Trix 2024, Zoho Corporation
Won Zoho's internal bot-building competition out of 70,000 participants. Built an HR workflow automation extension published to the Zoho Cliq marketplace. Awarded ā‚¹1,00,000 cash prize and internship offer.
1st / 70,000
3rd Place ā€” DravidianLangTech @ EACL 2024
Fake News Detection shared task leaderboard. Macro-F1: 0.84 across Dravidian languages.
3rd Globally
Top-10 ā€” FIRE 2023, Sarcasm Detection
5th and 8th place globally across two sarcasm detection tasks in Dravidian languages.
Top 10 Globally

Let's Connect

Get in Touch

madhumit007@e.ntu.edu.sg LinkedIn GitHub Google Scholar
References available upon request