Student. I build small tools — mostly machine learning, mostly to solve problems I have.
Projects
-
Antigravity Studio — Cloud agentic coding workspace for web and iPadOS backed by a pre-installed Vercel Gemini proxy, multi-turn tool loop (live web search, multi-repo GitHub inspection/commits, isolated JS VM + in-browser Pyodide Python), multimodal vision, and Studio artifacts. Live App.
-
Singapore Island Biodiversity — Remote-sensing ML pipeline predicting vegetation change on Singapore's micro-islands. Sentinel-2 NDVI + 22k iNaturalist observations → XGBoost, 33% better RMSE than persistence baseline. Live map · Writeup. Honest finding: current 10m satellite data can't yet resolve year-over-year change on islands this small.
-
JARVIS — Autonomous AI assistant with 60+ tools, persistent memory, RAG, browser automation, multi-agent workflows, voice interaction, and runtime tool generation.
-
donecheck — Adaptive verification skill for Claude Code and coding agents. Generates audience-calibrated sanity checks (plain, standard, expert) to verify AI blast radius before merging.
-
Weather-Predictor — Random Forest classifier trained on 105k real CPCB observations. Chronological split, persistence baseline, 67.1% accuracy vs 61.6% baseline.
-
JEE-Quiz-Studio — Zero-dependency quiz app with per-topic accuracy tracking. Built to make my own revision more targeted. Live App.
-
Assistive Stick — Arduino walking stick for visually impaired navigation. Ultrasonic obstacle detection with distance-scaled beep frequency. Team project — I wrote the firmware.
Writing
- What I Learned Trying to Predict Kanpur's Air Quality — Why 67% accuracy taught me more than 90% would have.
Find me
- Portfolio: happ3ns.github.io
- Email: akshatkharkwal39@gmail.com
