Murray Chen

Rewrote and extended an open-source teaching project into a config-driven multi-agent platform without LangChain (ReAct loop, tool calling, RAG, MCP JSON-RPC 2.0 client/server), backed by 177 tests at 93% coverage Cut wall-clock time 75% on parallel 4-agent workflows (simulated 200 ms model latency) by building a BSP super-step graph engine with asyncio fan-out/fan-in, typed channel merges, and pre-run BFS validation. Enabled cross-process resume and node-level retry with zero re-execution of completed nodes by checkpointing every super-step to file or Redis (TTL snapshots + sorted-set index); modeled human approval as a resumable interrupt.

Multi-Agent Platform & Workflow Engine (2026-10 – Present)

• Rewrote and extended an open-source teaching project into a config-driven multi-agent platform without LangChain (ReAct loop, tool calling, RAG, MCP JSON-RPC 2.0 client/server), backed by 177 tests at 93% coverage
• Cut wall-clock time 75% on parallel 4-agent workflows (simulated 200 ms model latency) by building a BSP super-step graph engine with asyncio fan-out/fan-in, typed channel merges, and pre-run BFS validation.
• Enabled cross-process resume and node-level retry with zero re-execution of completed nodes by checkpointing every super-step to file or Redis (TTL snapshots + sorted-set index); modeled human approval as a resumable interrupt.
• Achieved zero interrupted in-flight requests during agent hot-swaps with a reference-counted registry that swaps instances atomically and retires old ones only after every borrower releases them.

IR 3D Scanner with Autodiff Calibration (2026-09 – 2026-10)

• Cut 3D scan noise ~10× (4.6 → 0.42 cm per point) with zero added hardware by tracing it to servo holding current in a 25-second static experiment, then changing firmware to release servos and take a 7-reading median.
• Recovered 7.7× more usable scan points (119 → 917 of 1,125) by replacing a fixed 40–65 cm distance crop with filtering on the MLP’s training-voltage domain, rejecting extrapolation instead of the object itself.
• Made calibration 21× faster per gradient (30.7 → 1.45 ms) by writing pan-tilt kinematics in Taichi with reverse-mode autodiff (4e-9 error vs. finite differences) and fitting 7 error parameters with L-BFGS-B.
• Prevented silently wrong calibrations by adding an identifiability check (Gauss-Newton covariance + SVD null space) that locks undeterminable parameters and refuses to save non-converged or bound-hitting fits.
• Trained a voltage-to-distance MLP to 0.24 cm held-out MAE (R² = 0.9998) using distance-grouped hold-out to prevent leakage, and covered the pipeline with 23 hardware-free tests that run in ~5 s.

Clash Royale AI: Screen-Only Game Agent (2026-05 – 2026-05)

• Rebuilt a screen-only game agent as one PyTorch model (YOLO → Perceiver IO → Decision Transformer, ~2,850 lines), replacing a 5-component JAX/PaddleOCR/ResNet pipeline and removing ~6 GB of legacy code and data.
• Raised detection mAP50 4.6× (0.068 → 0.311) and recall 5× (5% → 27%) across 155 classes by building a data pipeline over 6,939 labeled frames (117K boxes) and fine-tuning YOLO26s on a 1-hour GPU budget.
• Cut per-frame attention cost from O(M²) to O(M·K) by designing a multimodal tokenizer that fuses 7 input types into one 256-d sequence, compressed through a 64-latent Perceiver bottleneck.
• Found and fixed 12 correctness bugs (target leakage into model inputs, padded frames counted in the loss, a class-ID collision) and locked them in with 9 regression tests that run in 1.4 s.

Bachelor of Science in Engineering in Electrical and Computer Engineering – Olin College of Engineering (2029-05)