Looks like you have JavaScript disabled. For the full Twine experience, you will need to re-enable it.

AI engineer who authors and verifies high-difficulty agentic benchmark tasks for frontier-model post-training and evaluation, owning environment design, programmatic verifiers, containerization, reward-hack hardening, and pass@k measurement end to end. Establishes task difficulty empirically against frontier reasoning models while also shipping production ML systems. MCA candidate (June 2026) with experience building reproducible benchmark environments and robust evaluation harnesses, plus production-ready ML platforms and autonomous multi-agent pipelines with streaming logs and provider fallback.…AI engineer who authors and verifies high-difficulty agentic benchmark tasks for frontier-model post-training and evaluation, owning environment design, programmatic verifiers, containerization, reward-hack hardening, and pass@k measurement end to end. Establishes task difficulty empirically against frontier reasoning models while also shipping production ML systems. MCA candidate (June 2026) with experience building reproducible benchmark environments and robust evaluation harnesses, plus production-ready ML platforms and autonomous multi-agent pipelines with streaming logs and provider fallback.

Chandrasekhar Thapa

AI Engineer, AI Evaluator, Programmer, +2





AI engineer who authors and verifies high-difficulty agentic benchmark tasks for frontier-model post-training and evaluation, owning environment design, programmatic verifiers, containerization, reward-hack hardening, and pass@k measurement end to end. Establishes task difficulty empirically against frontier reasoning models while also shipping production ML systems. MCA candidate (June 2026) with experience building reproducible benchmark environments and robust evaluation harnesses, plus production-ready ML platforms and autonomous multi-agent pipelines with streaming logs and provider fallback.…AI engineer who authors and verifies high-difficulty agentic benchmark tasks for frontier-model post-training and evaluation, owning environment design, programmatic verifiers, containerization, reward-hack hardening, and pass@k measurement end to end. Establishes task difficulty empirically against frontier reasoning models while also shipping production ML systems. MCA candidate (June 2026) with experience building reproducible benchmark environments and robust evaluation harnesses, plus production-ready ML platforms and autonomous multi-agent pipelines with streaming logs and provider fallback.

Available to hire

AI engineer who authors and verifies high-difficulty agentic benchmark tasks for frontier-model post-training and evaluation, owning environment design, programmatic verifiers, containerization, reward-hack hardening, and pass@k measurement end to end. Establishes task difficulty empirically against frontier reasoning models while also shipping production ML systems.

MCA candidate (June 2026) with experience building reproducible benchmark environments and robust evaluation harnesses, plus production-ready ML platforms and autonomous multi-agent pipelines with streaming logs and provider fallback.

Skills

Experience Level

Expert

Expert

Expert

Expert

Expert

Expert

Intermediate

Intermediate

Intermediate

Intermediate

Language

Work Experience

AI Engineer (Contract) — Agentic Benchmark & Evaluation Authoring at Handshake AI — Dynamo Program 2026

January 1, 2026 - Present

Author agentic-coding benchmark tasks for frontier-model post-training and evaluation. Each task is shipped as a reproducible Docker environment with an all-or-nothing reward contract and a programmatic verifier. Work spans debugging-and-repair, query optimization, and RL-infrastructure domains. Built containerized verification harnesses (including build integrity and containment), implemented reward-hacking batteries, and hardened solution-leakage vectors at the image and test-execution layers. Established empirical difficulty using measurement (trajectory/failure analysis, per-trap defeat probability modeling for pass@k forecasting, mutation testing to ensure requirements are load-bearing) and ran controlled disclosure A/B experiments to validate solve-rate improvements.