Projects

AI-generated courses built around what you already know, not a MOOC catalog or a chat tool. A short qualifying chat identifies your background and target gap, Claude drafts a tailored chapter plan, and each chapter is generated on demand with formatted text, typeset LaTeX math, and optional video, plus a persistent per-chapter Q&A chat. Built with my brother Sid.

One of two repos powering an automation pipeline that lets nontechnical archaeologists turn raw drone photos into 3D models and research sheets, built for the Tharros Archaeological Research Project in Sardinia, Italy. TARP Field runs on-site: a kanban-style dashboard tracks each photo job from capture through preliminary Metashape alignment and handoff to the lab. Job status pushes to Google Sheets every 5 minutes to stay in sync with TARP Lab, and misnamed job folders get flagged before anything goes missing.

The other half of the pipeline: TARP Lab picks up jobs handed off from TARP Field and turns them into finished 3D models and SU (stratigraphic unit) research sheets, all from a kanban dashboard, no command line or GIS experience required. Each job moves through 5 pipeline stages, and each unit through its own volumetrics pipeline (align → mesh → volume → SU sheet PDF). One click runs Metashape alignment, Poisson mesh reconstruction (CloudComPy), or georeferenced PDF generation (QGIS). Synced with Google Sheets so the Field and Lab machines share one source of truth.

A web app that stitches Google Street View screenshots into a full-width street elevation panorama, automating a manual step in urban planners' building inception reports. Idea came from conversations with urban planners in Mumbai and Boston. Tried 14 stitching approaches before landing on a PIL pipeline with luminance normalization and sigmoid cross-fade blending. Next.js frontend with drag-to-reorder image sequencing, FastAPI backend.

A from-scratch PyTorch implementation of I-JEPA (Assran et al., Meta AI 2023), a self-supervised vision model that learns by predicting in representation space instead of reconstructing pixels. ViT backbone with a context encoder, EMA target encoder, and a predictor. Cut training time by parallelizing embedding generation and using mixed precision. Currently exploring RoPE embeddings and the Muon optimizer for further speedups.

A browser extension that autofills job and accelerator applications using your profile and your own Claude API key. Add your profile once, fill any form in seconds, in Job mode (personal profile) or Accelerator mode (personal + startup profile). It scans every frame for fillable fields, including shadow DOM and cross-origin iframes, sends them to Claude with your profile as context, and writes the answers back using native input setters so React, Vue, and other frameworks register the change.

AI-powered pipeline that automates planset creation for urban bike lane infrastructure. Takes a street address and produces a 2-sheet PDF with gap analysis and plan drawings in under 40 seconds — collapsing what typically takes days into minutes. Uses NetworkX for topology-based connectivity gap detection, GeoPandas for map visualization, and Claude vision for generating engineering documents (cross-sections, construction notes, specs, quantity estimates). Built with my brother Sid.

A minimal iPhone habit tracker with a home screen widget. Track one habit, see your streak — that's it. Built with Swift and WidgetKit: a small widget shows the streak count, a medium widget has interactive check/X buttons via AppIntents. Daily notifications with inline actions, a configurable start date to preserve existing streaks, and all data stored in a shared App Group so the widget and app stay in sync.

Built in Sardinia, Italy as part of the Tharros Archaeological Research project. A script to automate volumetric generation for excavated stratigraphic units (SUs) using CloudCompare processes, producing 3D objects viewable from all angles for post-excavation analysis and record keeping. Collaborated with the University of Cincinnati. Output is archived on A.I.R., a visual archival platform for archaeological projects built by Paola Derudas at Lund University.

This site — a personal portfolio built with Next.js 13 (Pages Router), React 18, and Tailwind CSS. Deployed on Vercel. Inspired by Lee Robinson's website.

A concurrent Go server that brokers real-time WebRTC connections between browser clients and on-demand server clients, routing audio streams through Speech-to-Text. The backbone infrastructure for Ria Voice.

Voice agents for customer service — hacked together Vosk for ASR, a custom TTS pipeline (Tacotron / WaveGlow / HiFi-GAN) and Whisper for STT, all orchestrated by a concurrent Go server across multiple MediaStreams. Integrated GPT and tuned Whisper.cpp, cutting response time by 25%.

Interactive AI avatars whose faces move in sync with a voice agent's audio — think voice agent + a deepfake video model driven by speech. Designed and trained a custom TTS model, co-designed the lip-sync GAN. Pivoted to voice-only agents based on user feedback.

A map-based exploration of the artistic and geographical history of Pompeii. Built in collaboration with Dr. Eric Poehler (UMass Amherst) and Sebastian Heath (NYU) to visualize PALP's archaeological data.

A custom liquid-cooled compute server built for ML workflows: RTX 3070 (8GB vRAM), 128GB RAM, Ryzen 5 CPU. Hand-built the cooling loop with PEG hard tubes bent over candle heat — a $4K build during senior year finals that paid back in cloud compute savings.

A charitable donation platform that contextualizes contributions using cost-of-living data across countries, helping donors understand the real-world impact of their money.

An application that identifies objects in the real world through your camera and translates their names into your language — bridging the physical and linguistic gap for language learners.

A web app that uses your camera to capture hard-to-read text and reads it back to you aloud. Designed to be simple — ideal for the elderly, visually impaired, and those with reading difficulties.

A mobile app connecting users facing period emergencies with nearby community members willing to supply sanitary products — peer-to-peer, like a ride-share for menstrual supplies.

Helps users locate nearby trashcans via real-time mapping, and enables organizations to identify areas lacking waste disposal infrastructure.

A prototype educational VR headset built from a Jetson Nano, Google Cardboard, Unreal Engine and TensorRT. Developed hand-tracking neural nets for embedded inference (TFLite/LiteRT) to control the interface with hands. Built a whiteboard app with finger-drawing and educational content.

Built on Induino — a low-cost, Indian alternative to Arduino. An ultrasonic sensor fires sound waves and measures the distance to the nearest object by timing how long the echo takes to return, using the speed of sound to convert time into distance.