Things I've built

Classify waste images into 9 classes using the RealWaste dataset (4,752 images), as a college deep learning project.

Started with a baseline CNN, then fine-tuned MobileNetV2 using transfer learning with two-phase training. The winning ideas from the team's experiments were combined into the final configuration, with 5-crop test-time augmentation on top.

92.99% test accuracy and 0.9359 macro-F1, up from 69.7% for the baseline CNN and 84.3% for the early MobileNetV2. The weakest class is Miscellaneous Trash (F1 0.87). Presented and defended live.

Heavy projects were stacking up, and boilerplate plus copy-pasting code took a lot of my time. I first wanted to build my own LLM-based coding agent, but my hardware couldn't support it, so I shifted to features it could handle, like a parser and automations such as pasting code from the clipboard straight into files. That grew into a personal task-automation agent with a web UI, task queue, schedules and Telegram remote control, running 24/7 on my laptop as a background service. It made finishing my other projects easier and faster.

Flask backend with SSE streaming and a React/Vite UI. A tool-using agent loop rotates across Groq, Gemini and OpenRouter models, with 5 selectable gears (presets). A safety policy classifies every tool call as LOW, MEDIUM, HIGH or DENY under SAFE, ASSISTED or AUTONOMOUS modes, with approval cards. A SQLite queue and scheduler handle tasks.

v1 complete (finished around 11 Oct 2026).

A food donation app that connects donors and recipients. The base app was judged too basic, so an ML layer is being added to make it smarter about food safety and demand.

Started in Flutter, then rebuilt as a responsive React web app, with Capacitor wrapping it for mobile. Donations follow a 4-phase claim flow: Post, Request, Pickup, Delivered. A recipient's request can be cancelled while it's still pending. Planned, not built yet: a cooked-meal freshness classifier and donation-volume forecasting. A rule-based FSSAI safety layer comes first, a 50-60 image pilot dataset acts as a go/no-go gate before any CNN training, and Grad-CAM is planned for explainability.

In progress. The base app is built; the ML layer is still planned and not built yet.

Build a baseline road-damage detector (mainly potholes) with honestly measured scores, plus a small decision layer on a dashcam video demo. The limits are stated on purpose: not field-tested on Indian roads, not real time, and no accuracy claim on video.

Trained a YOLO11n baseline, then YOLO11s, on the RDD2022 India data (4 classes: longitudinal, transverse and alligator cracks, and potholes). The India-only model barely worked on UK pothole photos (Kent dataset), so I fine-tuned on India + Kent. A confidence cutoff of 0.35 was chosen on validation, because a wrong box seemed worse than a missed pothole. The video demo adds ByteTrack tracking, an editable road-area filter, duplicate-ID merging and a transparent 0-100 severity score.

YOLO11s on the India test set: mAP50 0.332 and pothole AP50 0.453. The default India + Kent model scores pothole AP50 0.413 on the India test and 0.218 on a held-out Kent test, up from 0.020 before fine-tuning. At the 0.35 cutoff about 4 in 10 potholes are found (recall 0.366 on India, 0.305 on Kent). Single runs, and the scores are slightly optimistic because the split comes from the Kaggle upload.