Founder & Machine Learning Engineer · Next AI
Founded the company and set its research direction, from post-training through low-latency serving. The platform now serves 1,000+ enterprise real estate clients across Southeast Asia.
A loop that mines its own failures
Automated verifiers mine production traces for model weaknesses, turn them into targeted eval suites, and feed critique-and-revise cycles back into training. Task success rose 18% over four rounds.
End-to-end real estate workflows
Designed and aligned agents with RLHF/RLAIF, tool use, retrieval, memory, and test-time planning. Multi-step completion up 23%; 12 client workflows automated with no human intervention.
The stack that gates every release
Extends SWE-bench, τ-bench, and LM Evaluation Harness with domain benchmarks for planning depth, tool-invocation precision, step-level hallucination, and multi-step reliability. Catches 87% of regressions.
Distributed training and inference
Engineered training and serving across TPU and GPU clusters in JAX and PyTorch FSDP. System-level profiling and custom CUDA kernels cut training time 40%.
