AI/ML Engineer

Nemotron-3-Ultra-550B, Gemma4-31B on 8xA100 40GB GPUs as production AWS SageMaker endpoint with tensor-parallel inference, Gemma4-12B-it on Lambda GPU, Megatron-Bridge w vLLM or SGLang,

▎ Fleet-evaluation harness for security-agent benchmarks 𝗖𝘆𝗯𝗲𝗿𝗚𝘆𝗺-𝗘𝟮𝗘, 𝗘𝘅𝗽𝗹𝗼𝗶𝘁𝗚𝘆𝗺, 𝗘𝘅𝗽𝗹𝗼𝗶𝘁𝗕𝗲𝗻𝗰𝗵 𝗩𝟴, 𝗖𝗩𝗘-𝗯𝗲𝗻𝗰𝗵, run in real disposable Daytona sandboxes per trial — with Modal-based KVM-capable sandboxes planned for kernel-exploit tasks. Measures agent capability and infra reliability, not simulated.

HoloSkill Gym: framework for self-evolving CLI coding-agent skills using Holo-CUA skill mutations with SkillOpt’s acceptance gate. SEAGym independently evaluates and checkpoints each evolving skill state. Harbor provides isolated, containerized execution for agent rollouts and verification (Skill evolution & evaluation transfer).

RL agents across LangChain & LangGraph, 50+ MCP registry tools, PGVector stores, deployment w Dynamo and inference with vLLM & SGLang

Inductive Transfer learning using Deep Convolutional Neural Networks (DCNNs) for fine-grained Knife classification using Metropolitan Police Data (knife stabbings incidents)

Hierarchical improvement flywheel for agents using OAI SDK with real traces, add human and model feedback, turn that feedback into evals, and use the resulting evidence to propose the next harness changes for Codex to implement.