Anh Dao

Teleport Contest -- Autonomous Coding Harness

• Participating in the Teleport Coding Challenge, using an agent swarm to port NetHack 5.0, a complex roguelike game with 500,000 lines of C and Lua code, to JavaScript with byte-exact behavior.
• Designing guardrails, multi-agent orchestration, evaluation mechanisms to prevent common pitfalls with independent long-horizon agents like reward hacking, challenge bypass and goal misalignment.

Personal Assistant on Cloudflare Durable Objects

• Building a personal assistant on Cloudflare Durable Objects with the Pi agent harness and OptChat-based memory: a never-ending chat compressed into a summary tree, giving constant-size context the agent can zoom into.

MLX Silicon Inference assorted projects

• Ported Open Source Decision models into an MLX-compatible framework, including Cloudflare Clef. Evaluated and optimized inference performance on an M4 Pro with relevant Jev-based evals and benchmarks.
• In separate llama.cpp experiments using Metal, speculative decoding increased Gemma 4 12B throughput by 12.1% across three tasks.

Scalable Agentic Systems Evaluation

• Ran SWE-Bench Lite evaluations on 300 software engineering tasks, utilizing open source models like GLM-4.7, across different harnesses like mini-swe-agent and Claude Code
• Deployed vLLM with ROCm on LUMI's AMD MI250X GPUs using SLURM; profiled GPU utilization, memory, latency, and concurrency for 6+ open-weight LLMs on these same tasks, fixing incompatible inference formats and missing support for AMD-based inference cookbooks.

GraphRAG with LRU Caching

Graph-based RAG pipeline using Kuzu and DSPy with LRU caching. Achieved 60-90% memory savings with Text2Cypher query refinement.

Encrypted File Transfer Protocol

Custom TCP protocol with AES-GCM encryption and RSA key exchange using pycryptodome. Defends against eavesdropping, tampering, and replay attacks.

Parallel Sudoku Solver

GPU-accelerated solver using CUDA for parallel computing on 100+ boards simultaneously. Implements constraint propagation optimized for GPU execution.