research.json

Experiments & findings

A running log of things I poke at - LLMs playing games, vision models driving phones, and benchmarks that probably shouldn't exist. Each entry ships with a full report.

/ live index

Browse the full research site

research.pythonicvarun.me

Open
/ filter
R/09

LLM Minesweeper Benchmark

Can Gemini 2.0 Flash play Minesweeper on a real Android device via ADB and uiautomator2? Step-by-step game replays with findings on vision-driven mobile automation.

full report →
LLMBenchmarkComputer-VisionAndroid
R/08

LLM Chess Evaluation Harness

A sophisticated evaluation framework for testing the chess-playing capabilities of Large Language Models (LLMs) against Stockfish. Uses MCP server and TUI dashboard.

full report →
LLMBenchmarkStockfishMCP-Server
R/07

Environmental Change Insights

A curated collection of major ecological disruptions and urban terraforming events detected across our study regions using OlmoEarth satellite data.

full report →
Geospatial-AIOlmoEarthSatellite-ImageryChange-Detection
R/06

Circuit Lab - LLM × Wokwi

A purely-browser playground that lets any OpenAI-compatible LLM generate complete Wokwi simulation projects from natural-language prompts, and runs them live in an embedded Wokwi simulator.

full report →
LLMLLM-Code-GenWokwi-SimulationEmbedded-Systems
R/05

Motion.dev MCP Evaluation

A comparative evaluation framework assessing the speed, implementation accuracy, and prompting strategies of AI agents (Claude Code, Codex, Gemini CLI, OpenCode Minimax) building interactive UI animations using the Model Context Protocol (MCP) and Motion.dev.

full report →
ReactViteTypeScriptMotion.dev
R/04

Remote Coding Agents

A self-hosted web workbench featuring a live file tree, terminal, and chat panel for running, orchestrating, and observing containerized AI coding agents inside isolated Docker sandboxes.

full report →
ReactTypeScriptNode.jsDocker
R/03

ASR Transcription Benchmark

Multi-model Automatic Speech Recognition (ASR) benchmarking suite testing OpenAI Whisper Realtime, Gemma, and Gemini models across clean baseline, noise, multi-speaker, and accent variations.

full report →
LLMBenchmarkASR-EvaluationWhisper-Realtime
R/02

Python Libraries Analysis

Static code quality audit of the most-downloaded PyPI libraries. Analyzes severity mapping using ty (Astral) and visualizes results via interactive error-density treemaps.

full report →
PythonStatic-AnalysisAST-ParsingClickHouse
R/01

AI-Powered Textbook Fact-Checking

An agentic fact-checking system designed to verify textbooks and technical documentation. Utilizes multi-agent extraction, deep web search, and structured verification reports.

full report →
LLMAgentic-RAGMulti-Agent-SystemsClaim-Extraction