Robotics & Automation engineer who moved into AI systems. I build the layer between LLMs and reliable production behavior: execution sandboxes, orchestration graphs, and evaluation harnesses that make agents debuggable instead of magical.

Benchmarked the cold-boot cost of three isolation backends for multi-tenant execution — Docker (runc, shared kernel), self-hosted Firecracker via kata-fc on Kubernetes (k3s · devmapper · RuntimeClass), and managed Firecracker via E2B — with measured p50/p95/p99 across two layers, per-backend security analysis, and a decision matrix. Finding: the microVM isolation boundary itself costs only ~0.8–1.0s over a container — most of the visible 2.5–4.3s spawn gap is JupyterHub startup and orchestration, not the kernel boundary. Measured, not guessed.

• Agent execution infrastructure @ LastLab AI: hybrid sandboxed runtime with conditional routing between Kata Containers (GPU) and Firecracker microVMs (CPU-only isolation) — the isolation stack I benchmarked publicly above
• LangGraph orchestration: stateful multi-step agent reasoning with typed tool contracts and recoverable execution graphs
• Eval frameworks: component-level golden-response testing for agent reliability

• Research paper implementations (PyTorch): BitNet 1-bit LLMs · Byte Latent Transformer · Multi-Query Attention · Milvus vector DB
• LangGraph multi-agent systems: self-correcting net-zero planner · real-time drone pest-detection (RT-DETR + SegFormer)

Agent reliability • Sandboxed execution • Specification-first systems design • Eval-driven development • The intersection of AI infrastructure and the physical world

B.Tech, Robotics & Automation (9.1 GPA) · Gemini/Llama training & evaluation pipelines @ Turing · Layer-1 protocol engineering in Go @ Cubane · robotics & UAV systems background

Firecracker microVMs • kata-containers / gVisor • LangGraph • Agent evaluation • Sandbox isolation tradeoffs • FastAPI • UAV systems

Email: sakshambisen123@gmail.com
LinkedIn: linkedin.com/in/sbsaksham