Focused on building intelligent systems that create real-world impact. I'm an MEng AI student at Duke University with experience in ML engineering, DevOps, and AI product development. I enjoy turning research ideas into practical solutions — whether that's a safety-critical CV system for search and rescue, a bias audit for transportation AI, or a multimodal fashion discovery engine. I bring a strong problem-solving mindset and love working at the intersection of AI and social good.
Duke Trust Lab
Jan 2026 — Present
BMW Group
Jan 2026 — Present
Duke University
Jan 2026 — Present
Assetmantle
Sep 2023 — May 2025
Hewlett Packard Enterprise
Jan 2023 — Jul 2023
95%
Python
85%
PyTorch
82%
TensorFlow
88%
FastAPI
85%
Docker
80%
Kubernetes
90%
AWS
75%
Google Cloud
85%
SQL
90%
Git
75%
Jenkins
85%
Linux
90%
Pandas
90%
NumPy
78%
JavaScript
72%
TypeScript
70%
React
68%
Next.js
92%
Jupyter
80%
MLflow
35 projects · hover to explore
Trained a lightweight RL policy network for multi-agent dialogue intervention using PyTorch and NLP. Reduced unnecessary interventions by 25% while maintaining task success rate, validated on a 10,000-dialogue simulation suite via A/B testing. Co-authoring a NeurIPS submission on safe agentic AI behavior under faculty supervision at Duke Trust Lab.
Fine-tuned Faster R-CNN on thermal SAR imagery with domain-specific augmentations (snow, smoke, sensor noise). Achieved 20% recall improvement under adverse conditions with <5% clean-data accuracy loss. Built end-to-end CV pipeline in PyTorch with custom data loaders and metrics optimized for safety-critical deployment.
Built and evaluated ML regression models (LightGBM, Ridge, MLP, Optuna-tuned TabularMLP) on real BMW dealer sales data to predict inventory turnover speed. Delivered explainable AI-driven inventory recommendations through a live REST API and interactive dealer dashboard, enabling data-informed spec decisions without requiring ML expertise.
Audited transportation AI allocation algorithms for racial and economic bias across Durham, NC — documenting that Black residents (32% of population) account for 47% of pedestrian/cyclist crash victims, driven by systematic undercounting of demand in low-income and high-minority census tracts. Built an interactive Leaflet mapping platform comparing AI-predicted vs. need-based infrastructure allocation using Census, NCDOT crash records, and OpenStreetMap data.
Built a reflective AI journaling companion using RAG over 8 psychological frameworks (CBT, IFS, NVC, Attachment Theory) to pattern-match emotional entries. Implemented cognitive distortion profiling across 4 distortion types and weekly self-awareness reports. Deployed full-stack Python/FastAPI on Railway with session persistence, mood trend tracking (1–10 daily score), and a trigger map logging recurring emotional patterns.
Built a film-to-fashion identification platform using FashionCLIP (512-dim embeddings) and FAISS GPU vector search over 20,000 Fashionpedia garment crops across 46 categories. Four-stage recommendation pipeline (FAISS → NeuMF → SASRec → diversity filter) with measurable NDCG@10 gains over a popularity baseline. Deployed FastAPI on Railway, Next.js on Vercel; evaluated on Precision@K, Recall@K, and MAP@K with 500 synthetic users × 30 interactions.
Built a SARIMAX inflation forecasting pipeline combining port traffic indicators with CPI data, reducing forecast error to 0.67–1.69% MAPE across major CPI categories. Validated model stability via 5-fold rolling cross-validation, achieving reliable predictive lift over CPI-only baselines.
Built a real-time, privacy-first Chrome extension estimating energy, carbon, and water footprint for AI prompts using model metadata and GitHub Models API. Implemented live footprint labels, heuristic + AI prompt optimization, and a client-side dashboard with daily impact summaries.
Built an EU AI Act compliance assessment system using TF-IDF + Logistic Regression, rule-based article evaluation (Articles 5, 6, 9, 10, 14), and automated remediation planning. Deployed full-stack ML workflow with MLflow, FastAPI, Docker, and Streamlit on Google Cloud Run, enabling explainable risk scoring and documentation audits.
Developed a semantic search engine for legal documents using embedding-based retrieval and citation-link modeling to surface semantically similar case law and statutes. Implemented vector search pipelines with passage ranking to support explainable legal research and faster knowledge discovery.
Edie is an AI cursor that fills a clinical form with you while you talk to your patient. Page and demo video: https://shreya-mendi.github.io/edie-cursor/ - Edie's cursor (light blue) follows the clinician-patient conversation. She moves to the question being discussed and leaves a suggested answer on it, with the exact words she heard.
How did approved gene therapies prove it, and what will FDA ask of you? 28 licensed gene therapies · 235 FDA documents · 5,766 pages · every answer cites a page number You are designing the safety and efficacy package for a gene therapy. You want to know what the approved products in your class actually did, and what FDA made them do
🔗 Live site: shreya-mendi.github.io/Website-Portfolio Personal portfolio site showcasing my projects and experience as an AI engineer. Built with plain HTML, CSS, and JavaScript, hosted on GitHub Pages.
🔗 Live app: youandme-xi.vercel.app Meals, workouts & dates — together. u-n-me is a private, two-person shared life app for a couple: a weekly meal plan, a collaborative grocery note, a workout planner, a date-night planner, a lightweight nudge feed, a recipe editor, and real recommenders that learn what the household likes. It has a tiny serverless backend (Vercel
🔗 Live site: shreya-mendi.github.io/UnmaskingReflections/ My reflections on the Unmasking AI book
🔗 Live site: shreya-mendi.github.io/mindguard-benchmark/ > Content Warning: This repository contains synthetic prompts simulating mental health crises at various severity levels. The content is designed for AI safety evaluation research only. If you or someone you know is in crisis, please contact the 988 Suicide & Crisis Lifeline (call or text 988) or the Crisis Text Line (text HOME to 741741).
🔗 Live demo: shreya-mendi.github.io/epstein-paper-trail/ > "The Documents Don't Lie" Paper Trail is an NLP system that classifies legal consequences for individuals named in the Epstein case using entirely public documents. It features a RAG-based chatbot, a named entity recognition (NER) pipeline, consequence classification across four tiers, and a dark interactive timeline UI.
🔗 Live demo: shreya-mendi.github.io/Tradecraft/ > Five specialized AI agents that research, signal, risk-check, execute, and audit trades — with a polished dark-luxury dashboard. ├── server.py # FastAPI backend (Phase 2) ├── orchestrator.py # CLI pipeline runner (Phase 1) ├── requirements.txt
🔗 Live demo: shreya-mendi.github.io/Catanist/ An instrumented arena where LLM players compete at Settlers of Catan — the sister project to the Mafia Arena. Different models (via GitHub Models) sit around one board, each primed with a controlled intent (diplomatic, cutthroat, deceptive, greedy, …), and you spectate the match in a cute illustrated replay
🔗 Live demo: shreya-mendi.github.io/Mafia-arena/ An instrumented arena where LLM players play social-deduction Mafia. Unlike a win-rate leaderboard, this treats persona, intent (cruel, peace-mongering, deceptive, truth-seeking, …), group size, and composition as controlled knobs — then records everything and renders a watchable replay.
A friendlier version of the CSV: each labeller reads a model's reasoning trace and picks which kind of intervention the reasoning claims to make (6 labels). The rubric is always on-screen, progress autosaves, and it's blind to the GPT-5 judge. pip install -r requirements.txt streamlit run app.py Opens at http://localhost:8501. Each person enters a name; their labels save to
Multimodal, faithfulness-grounded patent infringement discovery. Given a U.S. patent, PAEwall ranks the companies most likely infringing it, generates a per-limitation claim chart scored against real product documentation, and co-generates red-team non-infringement and invalidity arguments — giving patent holders and defendants the same honest view of enforcement viability.
This project builds a 3D interactive eco-city simulation where a Reinforcement Learning (RL) agent learns to design and manage a sustainable city over time. Unlike static simulations, the agent: - makes sequential urban planning decisions - balances competing objectives (growth vs sustainability) - adapts to dynamic environmental feedback
AIPI 590 | Midterm Project Platform: Raspberry Pi 4 | Language: Python 3 Pool Cue Assist teaches correct billiards stroke mechanics by detecting and classifying cue stroke quality in real time. The goal is stroke straightness and smoothness — the most common cause of missed shots is cue twist and lateral wobble on the follow-through, not aim. The device uses an IMU mounted on the cue to measure ho
> MEng AI Final Project — Explainable Artificial Intelligence > A multimodal agentic AI that plays Wordle by fusing a vision model and an > information-theoretic NLP solver, then surfaces real-time explanations of > why it made each decision — bridging the gap between raw model predictions > and human-understandable reasoning.
QuietSky is a calming, encouragement-focused speech adventure game designed to help players practice fluent, confident speech through fun gameplay—not through clinical or corrective cues The game contains four distinct modes, each modeled after genres the player enjoys (Flappy Bird, light exploration, light combat, and cockpit/FPS HUDs). Each mode supports a different aspect of fluent speech behav
We used ChatGPT 5 on 11/11 at 12:45pm to generate some code and content for this project. This repo contains a compact machine-learning pipeline that predicts whether a Disney hero headlines a blockbuster (above-median inflation-adjusted gross) by blending: - Character descriptors (hero, villain, signature song) from data/disney-characters.csv.
Modelling the ATM data for alternate data class For the data.ipynb script to work.Select all files in the atm folder and Download the data from duke.box.com . It will automatically be a .zip file. The python notebook is optimized to use the .zip folder and extract all the datasets into dataframes.
This project analyzes the World Happiness Report data to explore how happiness scores evolve across countries and regions over time (2019–2024). We investigate the role of key explanatory factors such as GDP, social support, healthy life expectancy, freedom of choice, generosity, and perceptions of corruption.
