CUHK (computer engineering) · Founder of NeuralZoo · Contributor to ezkl with a merged/ongoing PR for optional custom PWL lookup tables (Sigmoid replacement, input-range enforcement, soundness checks).
Focus: verifiable inference, ZKML circuits, custom lookup tables, privacy-preserving inference, AI agents.
| 🔬 | Secure Tensor Inference Protocol — private LLM inference; server only sees permuted tensors. (STIP-MLX, Apple Silicon / MLX) |
By underlying logic: Circuit (ZK / lookup), Agent (structured flows), Simulation (privacy / inference).
| Logic | Project | Core technical challenge |
|---|---|---|
| Circuit | ezkl | Pluggable PWL lookup in Halo2 circuits; JSON-defined table replaces built-in Sigmoid with input-range and soundness checks. |
| Circuit | ezkl-custom-lookup-experiment | Quantile-based PWL breakpoints and full prove/verify pipeline; built-in vs custom lookup compared end-to-end. |
| Circuit | zkml_lookup_experiment | Sigmoid lookup vs PLA: circuit size vs accuracy trade-off and metrics (e.g. 1–2 orders of magnitude with custom PWL). |
| Simulation | STIP-MLX | Block-diagonal permutation so the server only sees permuted tensors; private LLM inference on Apple Silicon with MLX. |
| Agent | relationship-consultant-agent | Agent pipeline for structured consultation flows. |
| Language | Zero-Knowledge | AI / ML | Robotics |
|---|---|---|---|
| Rust · TypeScript · Python | Halo2 · ezkl · PWL lookup · zk-SNARKs | PyTorch · ONNX · MLX | — |