AI Research Engineer

The core responsibilities for the job include the following:

Synthetic data generation:

• Design, extend, and evaluate our agentic pipeline for generating domain-specific pre-training and instruction-tuning data from public datasheets, reference manuals, and open-source embedded codebases.
• Invent and test data generation strategies: specification-to-code synthesis, compliance annotation, formal requirement extraction from natural language, and multi-step reasoning trace generation from hardware documentation.
• Evaluate data quality rigorously, not just statistical measures, but whether models trained on the data actually improve on real embedded engineering tasks
• Identify the highest-value data gaps in our training corpus and design generation pipelines to fill them.

Domain-specific model training and adaptation:

• Own continued pre-training and instruction-tuning runs across H2Loop's domain-specific model families.
• Design and evaluate training recipes: data mixture, tokenizer configuration, instruction format, curriculum, and RLHF/RLAIF alignment approaches.
• Benchmark model families rigorously not just on perplexity, but on task-level accuracy on hardware-specific code generation, compliance repair, and specification-grounded reasoning.
• Maintain H2Loop's model evaluation infrastructure: curated benchmark suites, regression pipelines, and human eval protocols tied to real customer tasks.

RL with hardware feedback:

• Design and run reinforcement learning experiments using real hardware boards as the reward environment, generating code that either works on the hardware or doesn't, producing a ground-truth training signal that no synthetic benchmark can replicate.
• Develop reward models and preference data pipelines from hardware pass/fail signals, user feedback, and formal verification outcomes.
• Investigate and prototype sample-efficient RL approaches suitable for the low-throughput, high-cost signal that physical hardware evaluation provides.

Neurosymbolic methods and formal verification:

• Research and prototype approaches that combine neural code generation with symbolic reasoning and formal analysis tools.
• Investigate feedback loops between generative models and verification systems and how verification outcomes can improve model behavior over time.
• Explore training techniques that make model-generated code more amenable to formal analysis without requiring explicit instruction at inference time.

Research translation:

• Monitor the research landscape across the areas relevant to H2Loop's stack: code generation, program synthesis, neurosymbolic AI, continual learning, RL for code, formal verification, and domain adaptation.
• Run experiments to evaluate whether promising techniques hold up on embedded/systems tasks; many results from general coding benchmarks do not transfer.
• Produce clear findings that drive product and model decisions: what to adopt, what to discard, and what to invest in further.

First 90 days:

• Days 1-30 Get deep on H2Loop's AI stack, our data generation pipelines, model families, RL environments, and verification tooling. Run existing training and evaluation pipelines end-to-end. Form a clear view of where the biggest research leverage is.
• Days 30-60 Run a focused experiment: a new data generation strategy, a training recipe improvement, a formal verification repair loop prototype, or an RL reward model evaluation. Produce findings with clear implications for the roadmap.
• Days 60-90 Propose a research agenda for the next two quarters. Own at least one research thread end-to-end from experimental design through evaluation to a concrete product or model outcome.

Requirements:

• PhD or equivalent research experience in machine learning, NLP, or a closely related field, or 4+ years of industry research with a publication record you can defend.
• Hands-on experience training or fine-tuning large language models: you have run training jobs, debugged training instabilities, and evaluated results against real task benchmarks, not just held-out loss.
• Strong foundations in deep learning and the transformer architecture: you understand what is happening during pre-training, instruction tuning, and RLHF, not just how to call the APIs.
• Rigorous empirical methodology: you design controlled experiments, track what changes between runs, and resist overclaiming from noisy results.
• Strong Python engineering skills: you can implement ideas cleanly, build evaluation pipelines, and productionize experiments without needing a separate engineering team to translate your notebooks.

Strong-to-have:

• Domain knowledge in formal methods or programming verification: familiarity with model checkers (CBMC, Frama-C), theorem provers (Lean 4 Coq, Isabelle), or SMT solvers (Z3).
• Experience with reinforcement learning from human feedback (RLHF), AI feedback (RLAIF), or execution-based reward (RL from compiler/test/verifier outcomes).
• Knowledge of embedded or systems software: C/C++, RTOS, safety standards (MISRA, AUTOSAR, IEC 61508 DO-178C), hardware abstraction layers, or MCU architecture.
• Experience with synthetic data generation for language model training, not just data augmentation, but designing generation pipelines that produce novel, high-quality training signals.
• Published work on code generation, program synthesis, neurosymbolic methods, or domain adaptation for LLMs.
• Familiarity with the industrial deployment constraints of our customers: air-gapped environments, on-prem inference, and compute-constrained hardware.