About H2LooP

H2LooP.ai is building next-generation coding models for system engineers across embedded, robotics, and industrial IoT. Our core team comes from Google, Toshiba, Cisco, Bosch, Philips, and NXP. We are building the infrastructure that allows AI to reason about real hardware.

⭐ What You’ll Work On

Agentic Workflow Design: Implement and refine Re-Act (Reasoning and Acting) loops and multi-step agentic chains to solve complex engineering tasks like "Fix this linker error" or "Generate a BSP for this SoC."
Inference Optimization: Work on reducing TTFT (Time To First Token) using vLLM, TensorRT-LLM, or speculative decoding techniques to make the developer experience seamless.
IDE & Tooling Integration: Help build and maintain our VS Code / IDE extensions that serve as the primary interface for system engineers.
Backend & Model Serving: Optimize FastAPI/Python services that orchestrate model requests, context retrieval, and hardware-in-loop feedback.
System-Aware RAG: Build pipelines that allow agents to "read" local codebases and datasheets to provide context-aware suggestions.

🎯 Required Skills

AI & Agentic Frameworks: Hands-on experience with Hugging Face (Transformers/Diffusers), LangChain, or LlamaIndex.
Model Serving: Familiarity with vLLM, Text Generation Inference (TGI), or Ollama for local/server-side model execution.
Agentic Design: Understanding of Re-Act patterns, tool-calling, and autonomous agent architectures.
Inference Knowledge: Basic understanding of quantization (GGUF/NVFP4/AWQ) and how to optimize model throughput.

🌟 Good to Have

Exposure to Vector Databases (Chroma, Pinecone, or Qdrant) for RAG workflows.
Familiarity with Docker and GPU orchestration (NVIDIA-Container-Toolkit).
Interest in how AI agents can interact with "non-text" artifacts like schematics or waveforms.

🔧 Example Tasks

Implement a Re-Act agent that can look up a function in a local C-header file and use it to fix a compilation bug.
Benchmarking and deploying a quantized model using vLLM to hit sub-100ms latency targets.
Creating a feedback loop where model-generated code is automatically checked against a compiler, and errors are fed back to the agent for "self-correction."

🚀 What We Offer

Direct Ownership: Work on the core "bridge" between our research models and real-world users.
Elite Pedigree: Mentorship from a team that has shipped systems at Google, Bosch, and NXP.
Deep-Tech Exposure: Move beyond "wrapper" apps and dive into real AI infrastructure and inference engineering.
Competitive Stipend: Plus a high-growth environment with a clear path to a Founding Team role.

Applied AI / Full-Stack Intern