I'm a Machine Learning Engineer at Optimizely, working on LLM evaluation and routing, fine-tuning (QLoRA), and production ML pipelines on Argo Workflows and EKS. Recent work includes a benchmark of Optimizely's AI desktop app, Mark (formerly Opal), against third-party AI tools, and an on-device LiteLLM routing proxy that sends each request to a model sized for the task.
Before that I built backend services for Zeda, a US clinical trial management system used by several US clinics, at Brick Line Technology, and shipped AgroGPT, a RAG platform for precision agriculture, and a face-recognition attendance system at ACI PLC.
Outside work I build and measure LLM systems from the ground up, mostly around efficient inference and serving, model routing and inference-time compute.
- AxiomServe: an LLM serving runtime built from scratch, with batched prefill, manual decode, KV-cache reuse, continuous batching, request scheduling and streaming. Continuous batching cut p95 time-to-first-token to 119 ms, down from 503 ms with static batching.
- routewise: a cost-vs-accuracy study of LLM routing on real HELM results for 22 models from 2024 and 2025. At equal accuracy, routing cost 48% less than the best cheap 2024 model and 25% less than the best 2025 model, and beat every single model on accuracy.
- VeriSearch: verifier-guided LLM reasoning under fixed token budgets. Adaptive early stopping used 35% fewer tokens than full reranking at about 1 point lower accuracy.
Languages:
LLM / ML:
MLOps & Data:
Backend & Infra:
