Ask Mani's Assistant
Hi, I'm Mani's assistant. Ask me anything about his work, projects, or background.

Projects

Vein Visualizer Master's Thesis

Real-time near-infrared vein-segmentation system for a low-cost, portable IV-access assist device — targeting under ₹50,000 versus ₹500,000+ commercial vein finders, with a screen-overlay design that needs no calibration. Built on DeepLabV3+ with a MobileNetV3-Large backbone and an attention-gated decoder, benchmarked against 3 other architectures over 50 epochs. Reached the best validation score at roughly 1/6th the parameters of an Attention U-Net baseline, scored on IOT (Intersection-over-Target) — a task-specific coverage metric I defined because standard IoU over-penalises thin-line vein annotations for small width mismatches — reaching 0.78 IOT.

Supervisor: Prof. Hardik J. Pandya, IISc  |  Duration: Oct 2024 – Apr 2025

PyTorchDeepLabV3+MobileNetV3AttentionNIR Imaging
View Code
Yeast knowledge graph
Yeast Connectome — Knowledge Graph Bachelor's Thesis

An LLM pipeline that mined 84,427 publication abstracts to construct a yeast knowledge graph of 3.4M+ entities and several hundred thousand typed interactions across S. cerevisiae and S. pombe. Designed the extraction prompts, entity/relation schema and validation strategy. Published as first author in the Journal of Molecular Biology (2025).

Supervisor: Prof. Marek Mutwil, NTU Singapore  |  Duration: Aug – Dec 2023

PythonOpenAI APINLPKnowledge GraphsCytoscapeJS
Explore the Project
Visual taxonomy challenge
Visual Taxonomy — Product Attribute Prediction

Meesho Data Science Challenge (Kaggle). Predicted product attributes — colour, pattern, sleeve length — from catalog images using an EfficientNet-based model with a custom linear classifier and image augmentation, achieving F1 = 0.735. Reduces manual cataloguing time and improves catalog accuracy at e-commerce scale.

CNNEfficientNetPyTorchData Augmentation
RAG pipeline
Retrieval-Augmented Generation (RAG) Pipeline

A local-first RAG system over academic PDFs: Qdrant vector database, llama.cpp for local inference, and Sentence-Transformer embeddings — deployed as the chatbot backing this website. Demonstrates privacy-preserving, on-prem LLM deployment, relevant anywhere data can't leave the perimeter.

Qdrantllama.cppRAGNLP
View Code
Computational genomics eQTLs
Computational Genomics — eQTL Discovery

Research internship at the Centre for Brain Research (CBRAIN), IISc, under Prof. Shweta Ramdas. Analysed gene-expression data from smokers and non-smokers to discover novel expression quantitative trait loci (eQTLs), building a variant-calling and association pipeline with QTLtools and BCFtools over large genotype–expression datasets.

LinuxQTLtoolsBCFtoolsStatistical Genetics
Image classification and segmentation
Image Classification & Segmentation — ResNet + U-Net

End-to-end TensorFlow pipeline for binary image classification (ResNet feature extraction) and semantic segmentation (U-Net), with a custom-dataset preprocessing pipeline and evaluation via accuracy and Dice score.

TensorFlowResNetU-NetDice Score
View Code
Object detector for the visually impaired
Object Detector for the Visually Impaired

Portable assistive device pairing an ESP32-CAM with an ultrasonic sensor for object detection and distance estimation, with voice feedback delivered through a companion Kivy mobile app.

ESP32-CAMOpenCVKivyEmbedded ML
Algorithmic trading bot
Algorithmic Intraday Options Trading Bot

Automated intraday trading bot on the Upstox API using a 14-day moving-average strategy for trade decisions, with real-time execution and backtested validation.

PythonUpstox APIBacktesting.py
View Code
LSTM stock price prediction
LSTM-Based Stock Price Prediction

Time-series forecasting with an LSTM network: preprocessing and scaling of historical stock data, followed by a sequential architecture designed to capture temporal dependencies and forecast price movements.

TensorFlowLSTMTime Series
View Code