Projects

Autonomous vision-language agent that plays Survival 3D on YouTube Playables — screenshots the game, reads goals from the UI, plans movement, and executes keyboard actions in a loop.

• Closed-loop agent on YouTube Playables: capture screenshot → read goal, progress, and inventory → plan 1–2 compass moves → send key presses → persist memory for the next cycle.
• GoalMakerLLM reads the game UI and picks the fastest heading; PlannerLLM turns that into concrete NORTH/SOUTH/EAST/WEST steps — each component can use OpenAI or local Ollama vision models.
• MemoryManager keeps JSON knowledge of mechanics, discoveries, and failed strategies; RunLogger saves per-cycle screenshots and run.json under outputs/.
• Playwright + Chrome with a persistent browser profile for one-time YouTube login; uv-based Python 3.11+ setup with configurable GAME_URL and model providers.

Black-box UI defect discovery for locally hosted web apps — Playwright exploration, LLM-assisted goals and verification, and an HTML findings report.

• Four-phase pipeline: BFS exploration in a real browser → LLM converts paths into plain-English test goals → parallel goal-driven executors → per-step LLM verifier that emits structured findings.
• Executors resolve instructions at runtime instead of replaying raw trajectories — app-agnostic design with Docker lifecycle management, DOM heuristics, and crash-detection fallbacks when the LLM is off.
• Single CLI (run_harness.py) with composable modes — full pipeline, planner-only, run-goals, multi-pass rollout, and report-only — default target is a buggy Memos Docker image, but the harness generalizes to any local web app.

Skill and installer that produce a rich, self-contained HTML report explaining any Prime Intellect verifiers environment.

• Walks the environment source and emits environment_overview.html—dataset and task samples, every reward/metric with weights and scoring range, rollout logic from prompt to final score, and a full config reference.
• Ships as an npm CLI (npx understanding-prime-env) with interactive setup plus flags for Claude Code, Cursor, Windsurf, GitHub Copilot, Zed, or all tools in one go.
• Meant to be used from inside a verifiers env folder so you can ask your assistant to “explain this environment” and get an overview styled to match the Prime Intellect platform.

Ice-sliding multi-player maze environment on OpenEnv for benchmarking planning and coordination in RL agents.

• Ice-sliding mechanics — agents slide until hitting a wall or another player, requiring multi-step lookahead over reactive movement.
• Simultaneous multi-player movement; solved only when every player reaches an exit cell in the same phase.
• Reward shaping penalises repeated actions (−1), reversals, and revisited board states scaled by prior visit count.
• Standard OpenEnv reset()/step() API — compatible with any RL training loop without environment-specific wrappers.
• Includes dataset validation tooling and GIF rendering from recorded rollouts for debugging agent trajectories.

RL environment where an LLM agent negotiates prices against an LLM-powered seller using real marketplace listings.

• Three difficulty levels (Easy / Medium / Hard) with varying zones of possible agreement — from a $480 ZOPA down to no ZOPA, testing walk-away discipline.
• Six-component reward system on a [−1, 1] scale: surplus capture, walk-away correctness, output compliance, closing speed, opening offer quality, and concession smoothness.
• Walk-away penalty weighted 5× — trains the agent to value discipline over forcing a bad deal.
• Pure Python arithmetic grader ensures reproducibility; stochastic LLM sampling on both sides creates diverse training signal across episodes.
• Deployable on Hugging Face Spaces, Docker, or locally via uv; ships with a browser playground showing real-time reward breakdowns.

Full-stack interactive console for demonstrating and chatting with Reinforcement Learning models and agentic systems.

• Chat interface for real-time interaction with RL models, backed by smolagents for agent-based reasoning and step-by-step trace display.
• FastAPI backend (Python 3.12+) paired with a Next.js 15 / React 19 frontend — clean separation of model logic and UI.
• Integrates Hugging Face datasets directly; response caching via JSON files keeps repeated queries fast.
• Docker-first production deployment — single container exposes the full stack on port 7860, ready for Hugging Face Spaces.
• Environment-driven config (HF token, model name) makes swapping the underlying model a one-line change.

CrewAI-powered bot that auto-syncs multilingual documentation across feature branches via a Git pre-push hook.

• 2nd Runner-Up at the CrewAI Fall Agentic AI Challenge.
• Detects English documentation changes and auto-generates synchronised translations (Korean, Brazilian Portuguese) with no manual intervention.
• Runs as a Git pre-push hook — fires before any feature branch push, prompting the developer to sync docs before code lands.
• Agent-based architecture with specialised CrewAI agents for change detection, translation, and cross-branch synchronisation.
• Minimal setup: Python 3.12+, uv, and an OpenAI key — clone + uv sync + one hook registration.