Staff Machine Learning Engineer, LLM Fine-Tuning (Verilog/RTL Applications)

Level: Staff
Location: San Jose, CA
Cloud: AWS (primary: Bedrock SageMaker)

Why this role exists

You will architect and lead privacy-preserving LLM capabilities that support hardware design teams working with Verilog/SystemVerilog and RTL artifacts. This includes code generation, refactoring, lint explanation, constraint translation, and spec-to-RTL assistance. You ll lead a small, high-leverage team focused on fine-tuning and productizing LLMs in a strict enterprise data-privacy environment.

You do not need deep RTL expertise to start curiosity, LLM craftsmanship, and strong engineering rigor matter most. Exposure to HDL/EDA tooling is a plus.

Responsibilities

Technical Leadership & Roadmap

• Own the end-to-end roadmap for Verilog/RTL-focused LLM capabilities, covering model selection, fine-tuning, evals, deployment, and continuous improvement.

• Lead a hands-on team of applied ML engineers/scientists, unblock technically, review designs and code, and drive experimentation velocity and reliability.

Model Training & Customization

• Fine-tune and customize models using modern techniques (LoRA/QLoRA, PEFT, instruction tuning, RLAIF/preference optimization).

• Build HDL-aware evaluation workflows:

  • Compile/lint/simulate-based pass rates

  • Pass@k for code generation

  • Constrained decoding enforcing HDL syntax

  • Does-it-synthesize? checks

Privacy-First AWS ML Pipelines

• Design secure training & inference environments using AWS services such as:

  • Amazon Bedrock (incl. Anthropic models)

  • SageMaker or EKS KServe/Triton/DJL for bespoke training

• Implement strict privacy controls:

  • Artifacts in S3 with KMS CMKs

  • VPC-only infrastructure with PrivateLink (incl. Bedrock endpoints)

  • IAM least-privilege, CloudTrail auditing

  • Secrets Manager for credential handling

  • Full encryption in transit/at rest

  • No public egress for customer/RTL corpora

Inference & Deployment

• Stand up scalable, reliable LLM serving:

  • Bedrock model invocation where applicable

  • Low-latency self-hosted inference (vLLM/TensorRT-LLM)

  • Autoscaling and canary/blue-green rollouts

Evaluation Culture & Tooling

• Build automated regression suites running HDL compilers/simulators to measure correctness and detect hallucinations.

• Track experiments and produce model cards using MLflow/W&B.

Cross-Functional Collaboration

• Work with hardware design teams, CAD/EDA, Security, and Legal to:

  • Prepare/anonymize datasets

  • Define acceptance gates

  • Meet licensing, compliance, and security requirements

Productization

• Integrate models into engineering workflows: IDE plugins, CI bots, code review assistants, retrieval over internal HDL repos/specs, and safe function-calling.

Mentorship

• Develop team capabilities in LLM training, reproducibility, secure pipelines, and research literacy.

Minimum Qualifications

• 10 years total engineering experience; 5 years in ML/AI or large-scale distributed systems; 3 years with transformers/LLMs.

• Proven record shipping LLM-powered features and leading cross-functional technical initiatives at Staff level.

• Deep, hands-on experience with:

  • PyTorch, Hugging Face Transformers/PEFT/TRL

  • Distributed training (DeepSpeed/FSDP)

  • LoRA/QLoRA, grammar-guided decoding

• Strong AWS expertise:

  • Bedrock (model customization, Guardrails, Knowledge Bases, VPC endpoints)

  • SageMaker (Training/Inference/Pipelines)

  • S3, EC2/EKS/ECR, IAM, VPC, KMS, CloudWatch/CloudTrail, Step Functions, Secrets Manager

• Strong Python engineering fundamentals (testing, CI/CD, observability, performance tuning).

• Excellent technical communication and ability to set vision across teams.

Preferred Qualifications

• Familiarity with Verilog/SystemVerilog/RTL workflows (lint, simulation, synthesis, timing closure, test benches).

• Experience with static-analysis/AST-aware tokenization and grammar-constrained decoding.

• RAG over code/spec repos; tool-use/function-calling for code transformation.

• Inference optimization (TensorRT-LLM, KV-cache tuning, speculative decoding).

• Experience with enterprise model governance and security frameworks (SOC2/ISO 27001/NIST).

• Background in data anonymization, DLP scanning, and code de-identification.

What success looks like

90 Days

• Stand up HDL-aware eval harness with compile/simulate checks.

• Establish secure AWS training & inference environments (VPC-only, KMS encryption, no public egress).

• Deliver initial fine-tuned model with measurable performance gains.

180 Days

• Expand training coverage using Bedrock SageMaker/EKS.

• Add constrained decoding and retrieval over design specs.

• Productionize inference with SLOs and rollout to pilot teams.

12 Months

• Reduce RTL review/iteration cycles using measurable metrics: lint-clean time, defect reductions, suggestion acceptance rates.

• Establish a stable MLOps pathway for continuous improvements.

Security & Privacy by Design

• All sensitive data remains within private AWS VPCs with IAM-controlled access and CloudTrail auditing.

• Bedrock access via VPC PrivateLink endpoints only.

• Strict data minimization, tagging, retention, reproducibility, and DLP scanning.

• Model cards, lineage, and evaluation artifacts for each release.

Tech Stack

Modeling: PyTorch, HF Transformers/PEFT/TRL, DeepSpeed/FSDP, vLLM, TensorRT-LLM
AWS/MLOps: Bedrock, SageMaker, ECR, EKS/KServe/Triton, MLflow/W&B, Step Functions
Platform/Security: S3 KMS, IAM, VPC/PrivateLink, CloudWatch/CloudTrail, Secrets Manager
Bonus: HDL toolchains, vector stores (pgvector/OpenSearch), GitHub/GitLab CI