Shravani Hariprasad

Metrolinx GO Rail: Multi-Sensor Asset-Inspection Perception

An end-to-end camera + LiDAR + GPS/IMU pipeline for automated rail-asset inspection: a SLAM / visual-odometry branch localizes the vehicle in GPS-denied corridors while a VLM branch detects and condition-scores each asset; sensor-fused poses pin every detection onto a GIS map, with cross-pass change detection driving prioritized, geolocated work orders. 91% mAP across 14 asset classes, 0.7 m SLAM ATE over 18 km, 93% inspector agreement.

Applied · Hatch × Metrolinx

Camera-Only Train Localization in GPS-Denied Tunnels

An end-to-end monocular pipeline (Python, OpenCV; ORB tracking, essential-matrix pose estimation, RANSAC) estimating an MBTA Green Line train's position from forward cab video alone, ~1,200 frames/segment of 4K video on the Park Street to Government Center run. Domain-specific motion priors (motion gating, forward-motion constraint, rotation limits) improved trajectory consistency 0.41 to 0.79, with tracking sustained across 100% of the GPS-denied segment and station dwells auto-detected (31% of frames).

Implemented & validated · Hatch

CarlaPerception: Visual Perception & SLAM for Self-Driving

A from-scratch self-driving stack: YOLO detection, IoU tracking, DeepLabV3 segmentation, plus monocular to stereo visual odometry and loop-closure pose-graph SLAM, evaluated with ATE/RPE (Umeyama alignment) against KITTI ground truth: loop closure cut drift 62% (ATE 25 m to 9.5 m, 7 closures) and generalized to seq 05 (−33%) with no per-sequence tuning. VO validated against exact CARLA ground truth (ATE 0.59 m over 1,000 frames); includes a from-scratch GPU-trained monocular-depth network (AbsRel 0.21) with a sim-to-real evaluation, and COLMAP-free Gaussian-Splatting reconstructions (~641K Gaussians) from its own VO poses.

When Structure Is Hidden: LLM Evaluation Framework (First author)

First-author study introducing a two-pass evaluation framework and a conditional metric across 5 domains and 8 model configurations (GPT-4o, DeepSeek, Qwen2.5-7B, Llama-3.1-8B) to measure compositional reasoning failures in LLMs; 41 of 48 cross-domain comparisons statistically significant. Relevant to agentic-system and dataset evaluation.

Under review · EMNLP 2026

NYC Subway Ridership Pipeline & A/B Testing

End-to-end Airflow ETL ingesting MTA hourly-ridership open data into PostgreSQL/Supabase, processing 120M+ records across 428 station complexes with idempotent upserts and GitHub Actions checks. Difference-in-differences found a +6.4% ridership lift from the Grand Central Madison opening (p<0.001); a LightGBM forecaster reached 8.3% MAPE.

RAG Document Assistant

An AI-powered document assistant that answers questions over your files using retrieval-augmented generation with Gemini 2.5, served through a FastAPI + Streamlit interface.

TailTalk: AI-Powered Pet-Care Chatbot

Led a team building a RAG chatbot (React.js, Llama 2 via Hugging Face) with Pinecone vector search and low-latency semantic retrieval across 100+ pet-care and veterinary documents; deployed on AWS SageMaker.

Street-Level CV for Urban Change Analysis

HDMA Lab project curating a 10,000+ image dataset along downtown San Diego corridors (Google Street View, Mapillary), enriched with timestamped, geocoded metadata. A YOLOv8 detector with VLM enrichment identifies homelessness indicators (encampments, carts, makeshift shelters) at 88 to 92% mAP, and detections are compared across multi-year imagery of the same locations, visualized as ArcGIS point layers across downtown to surface block-level trends, with manual audits of flagged changes confirming ~90% change-detection precision.

PrEP-Related Messages Across Facebook, Instagram & Twitter

NIH-funded SDSURF study: collected 59K+ posts across Facebook, Instagram, and Twitter via CrowdTangle and the Twitter API and ran a quantitative content analysis on a 2,811-post random sample; found PrEP definitions/indications dominated content, information drew largely from non-traditional sources, and MSM were the most-mentioned population, informing HIV-prevention message design.

Under final review · JMIR

Interactive COVID-19 Data Visualization (CyberGIS)

HDMA Lab project building interactive spatial dashboards of San Diego COVID-19 data with a CyberGIS toolchain for exploratory epidemiological analysis.

Image Caption Generation (PySpark + GCP)

A deep-learning image-captioning pipeline trained at scale on Google Cloud with PySpark, generating natural-language descriptions for images.

Human Activity Recognition

Classifying human activities from sensor data with machine-learning and deep-learning models, with Tableau dashboards for performance and error analysis.

Public Sentiment on Climate Change

Analyzing public sentiment toward climate change to inform policy, using NLP, ML classification, and named-entity extraction, visualized in Tableau.

Cafe Management System

A relational database application for cafe operations, backed by MySQL on AWS RDS with Python tooling and normalized DBMS design.

Society Maintenance Management System

A full-stack web app for residential-society maintenance, built with React, Java Spring Boot, MongoDB, and REST services.

E-Zone: E-Commerce Platform (MERN)

A full e-commerce web application built on the MERN stack with product catalog, cart, and order flows.