SRS Consulting Inc
Staff Machine Learning Engineer- Local to CA
SRS Consulting Inc, San Jose, California, United States, 95199
Role:
Staff Machine Learning Engineer
Location: San Jose, CA (Onsite) Locals
Duration: Long-term
Why this role exists We’re building privacy preserving LLM capabilities that help hardware design teams reason over Verilog/SystemVerilog and RTL artifacts—code generation, refactoring, lint explanation, constraint translation, and spec‑to‑RTL assistance. We’re looking for a Staff level engineer to technically lead a small, high‑leverage team that fine‑tunes and productizes LLMs for these workflows in a strict enterprise data privacy environment.
You don’t need to be a Verilog/RTL expert to start; curiosity, drive, and deep LLM craftsmanship matter most. Any HDL/EDA fluency is a strong plus.
What you’ll do (Responsibilities)
Own the technical roadmap for Verilog/RTL focused LLM capabilities—from model selection and adaptation to evaluation, deployment, and continuous improvement.
Lead a hands‑on team of applied scientists/engineers: set direction, unblock technically, review designs/code, and raise the bar on experimentation velocity and reliability.
Fine‑tune and customize models using state‑of‑the‑art techniques (LoRA/QLoRA, PEFT, instruction tuning, preference optimization/RLAIF) with robust HDL‑specific evals:
Compile / lint / simulate based pass rates, pass@k for code generation, constrained decoding to enforce syntax, and “does it synthesize” checks.
Design privacy‑first ML pipelines on AWS:
Training/customization and hosting using Amazon Bedrock (including Anthropic models) where appropriate; SageMaker (or EKS + KServe/Triton/DJL) for bespoke training needs.
Artifacts in S3 with KMS CMKs; isolated VPC subnets & PrivateLink (including Bedrock VPC endpoints), IAM least privilege, CloudTrail auditing, and Secrets Manager for credentials.
Enforce encryption in transit/at rest, data minimization, no public egress for customer/RTL corpora.
Stand up dependable model serving: Bedrock model invocation where it fits, and/or low‑latency self‑hosted inference (vLLM/TensorRT LLM), autoscaling, and canary/blue‑green rollouts.
Build an evaluation culture: automatic regression suites that run HDL compilers/simulators, measure behavioral fidelity, and detect hallucinations/constraint violations; model cards and experiment tracking (MLflow/Weights & Biases).
Partner deeply with hardware design, CAD/EDA, Security, and Legal to source/prepare datasets (anonymization, redaction, licensing), define acceptance gates, and meet compliance requirements.
Drive productization: integrate LLMs with internal developer tools (IDEs/plug‑ins, code review bots, CI), retrieval (RAG) over internal HDL repos/specs, and safe tool use/function calling.
Mentor & uplevel: coach ICs on LLM best practices, reproducible training, critical paper reading, and building secure by default systems.
What you’ll bring (Minimum qualifications)
10+ years total engineering experience with 5+ years in ML/AI or large‑scale distributed systems; 3+ years working directly with transformers/LLMs.
Proven track record shipping LLM‑powered features in production and leading ambiguous, cross‑functional initiatives at Staff level.
Deep hands‑on skill with PyTorch, Hugging Face Transformers/PEFT/TRL, distributed training (DeepSpeed/FSDP), quantization‑aware fine tuning (LoRA/QLoRA), and constrained/grammar‑guided decoding.
AWS expertise to design and defend secure enterprise deployments, including:
Amazon Bedrock (model selection, Anthropic model usage, model customization, Guardrails, Knowledge Bases, Bedrock runtime APIs, VPC endpoints)
SageMaker (Training, Inference, Pipelines), S3, EC2/EKS/ECR, VPC/Subnets/Security Groups, IAM, KMS, PrivateLink, CloudWatch/CloudTrail, Step Functions, Batch, Secrets Manager.
Strong software engineering fundamentals: testing, CI/CD, observability, performance tuning; Python a must (bonus for Go/Java/C++).
Demonstrated ability to set technical vision and influence across teams; excellent written and verbal communication for execs and engineers.
Nice to have (Preferred qualifications)
Familiarity with Verilog/SystemVerilog/RTL workflows: lint, synthesis, timing closure, simulation, formal, test benches, and EDA tools (Synopsys/Cadence/Mentor).
Experience integrating static analysis/AST‑aware tokenization for code models or grammar‑constrained decoding.
RAG at scale over code/specs (vector stores, chunking strategies), tool use/function calling for code transformation.
Inference optimization: TensorRT LLM, KV cache optimization, speculative decoding; throughput/latency trade‑offs at batch and token levels.
Model governance/safety in the enterprise: model cards, red‑teaming, secure eval data handling; exposure to SOC2/ISO 27001/NIST frameworks.
Data anonymization, DLP scanning, and code de‑identification to protect IP.
What success looks like 90 days
– Baseline an HDL‑aware eval harness that compiles/simulates; establish secure AWS training & serving environments (VPC only, KMS backed, no public egress). Ship an initial fine‑tuned/customized model with measurable gains vs. base (e.g., +X% compile pass rate, Y% lint findings per K LOC generated).
180 days
– Expand customization/training coverage (Bedrock for managed FMs including Anthropic; SageMaker/EKS for bespoke/open models). Add constrained decoding + retrieval over internal design specs; productionize inference with SLOs (p95 latency, availability) and audited rollout to pilot hardware teams.
12 months
– Demonstrably reduce review/iteration cycles for RTL tasks with clear metrics (defect reduction, time to lint clean, % auto‑fix suggestions accepted), and a stable MLOps path for continuous improvement.
How we work (Security & privacy by design) Customer and internal design data remain within private AWS VPCs; access via IAM roles and audited by CloudTrail; all artifacts encrypted with KMS.
No public internet calls for sensitive workloads; Bedrock access via VPC interface endpoints/PrivateLink with endpoint policies; SageMaker and/or EKS run in private subnets.
Data pipelines enforce minimization, tagging, retention windows, and reproducibility; DLP scanning and redaction are first‑class steps.
We produce model cards, data lineage, and evaluation artifacts for every release.
Tech you’ll touch Modeling: PyTorch, HF Transformers/PEFT/TRL, DeepSpeed/FSDP, vLLM, TensorRT LLM.
AWS & MLOps: Amazon Bedrock (Anthropic and other FMs, Guardrails, Knowledge Bases, Runtime APIs), SageMaker (Training/Inference/Pipelines), MLflow/Weights & Biases, ECR, EKS/KServe/Triton, Step Functions.
Platform/Security: S3 + KMS, IAM, VPC/PrivateLink (incl. Bedrock), CloudWatch/CloudTrail, Secrets Manager.
Tooling (nice to have): HDL toolchains for compile/simulate/lint, vector stores (pgvector/OpenSearch), GitHub/GitLab CI.
#J-18808-Ljbffr
Location: San Jose, CA (Onsite) Locals
Duration: Long-term
Why this role exists We’re building privacy preserving LLM capabilities that help hardware design teams reason over Verilog/SystemVerilog and RTL artifacts—code generation, refactoring, lint explanation, constraint translation, and spec‑to‑RTL assistance. We’re looking for a Staff level engineer to technically lead a small, high‑leverage team that fine‑tunes and productizes LLMs for these workflows in a strict enterprise data privacy environment.
You don’t need to be a Verilog/RTL expert to start; curiosity, drive, and deep LLM craftsmanship matter most. Any HDL/EDA fluency is a strong plus.
What you’ll do (Responsibilities)
Own the technical roadmap for Verilog/RTL focused LLM capabilities—from model selection and adaptation to evaluation, deployment, and continuous improvement.
Lead a hands‑on team of applied scientists/engineers: set direction, unblock technically, review designs/code, and raise the bar on experimentation velocity and reliability.
Fine‑tune and customize models using state‑of‑the‑art techniques (LoRA/QLoRA, PEFT, instruction tuning, preference optimization/RLAIF) with robust HDL‑specific evals:
Compile / lint / simulate based pass rates, pass@k for code generation, constrained decoding to enforce syntax, and “does it synthesize” checks.
Design privacy‑first ML pipelines on AWS:
Training/customization and hosting using Amazon Bedrock (including Anthropic models) where appropriate; SageMaker (or EKS + KServe/Triton/DJL) for bespoke training needs.
Artifacts in S3 with KMS CMKs; isolated VPC subnets & PrivateLink (including Bedrock VPC endpoints), IAM least privilege, CloudTrail auditing, and Secrets Manager for credentials.
Enforce encryption in transit/at rest, data minimization, no public egress for customer/RTL corpora.
Stand up dependable model serving: Bedrock model invocation where it fits, and/or low‑latency self‑hosted inference (vLLM/TensorRT LLM), autoscaling, and canary/blue‑green rollouts.
Build an evaluation culture: automatic regression suites that run HDL compilers/simulators, measure behavioral fidelity, and detect hallucinations/constraint violations; model cards and experiment tracking (MLflow/Weights & Biases).
Partner deeply with hardware design, CAD/EDA, Security, and Legal to source/prepare datasets (anonymization, redaction, licensing), define acceptance gates, and meet compliance requirements.
Drive productization: integrate LLMs with internal developer tools (IDEs/plug‑ins, code review bots, CI), retrieval (RAG) over internal HDL repos/specs, and safe tool use/function calling.
Mentor & uplevel: coach ICs on LLM best practices, reproducible training, critical paper reading, and building secure by default systems.
What you’ll bring (Minimum qualifications)
10+ years total engineering experience with 5+ years in ML/AI or large‑scale distributed systems; 3+ years working directly with transformers/LLMs.
Proven track record shipping LLM‑powered features in production and leading ambiguous, cross‑functional initiatives at Staff level.
Deep hands‑on skill with PyTorch, Hugging Face Transformers/PEFT/TRL, distributed training (DeepSpeed/FSDP), quantization‑aware fine tuning (LoRA/QLoRA), and constrained/grammar‑guided decoding.
AWS expertise to design and defend secure enterprise deployments, including:
Amazon Bedrock (model selection, Anthropic model usage, model customization, Guardrails, Knowledge Bases, Bedrock runtime APIs, VPC endpoints)
SageMaker (Training, Inference, Pipelines), S3, EC2/EKS/ECR, VPC/Subnets/Security Groups, IAM, KMS, PrivateLink, CloudWatch/CloudTrail, Step Functions, Batch, Secrets Manager.
Strong software engineering fundamentals: testing, CI/CD, observability, performance tuning; Python a must (bonus for Go/Java/C++).
Demonstrated ability to set technical vision and influence across teams; excellent written and verbal communication for execs and engineers.
Nice to have (Preferred qualifications)
Familiarity with Verilog/SystemVerilog/RTL workflows: lint, synthesis, timing closure, simulation, formal, test benches, and EDA tools (Synopsys/Cadence/Mentor).
Experience integrating static analysis/AST‑aware tokenization for code models or grammar‑constrained decoding.
RAG at scale over code/specs (vector stores, chunking strategies), tool use/function calling for code transformation.
Inference optimization: TensorRT LLM, KV cache optimization, speculative decoding; throughput/latency trade‑offs at batch and token levels.
Model governance/safety in the enterprise: model cards, red‑teaming, secure eval data handling; exposure to SOC2/ISO 27001/NIST frameworks.
Data anonymization, DLP scanning, and code de‑identification to protect IP.
What success looks like 90 days
– Baseline an HDL‑aware eval harness that compiles/simulates; establish secure AWS training & serving environments (VPC only, KMS backed, no public egress). Ship an initial fine‑tuned/customized model with measurable gains vs. base (e.g., +X% compile pass rate, Y% lint findings per K LOC generated).
180 days
– Expand customization/training coverage (Bedrock for managed FMs including Anthropic; SageMaker/EKS for bespoke/open models). Add constrained decoding + retrieval over internal design specs; productionize inference with SLOs (p95 latency, availability) and audited rollout to pilot hardware teams.
12 months
– Demonstrably reduce review/iteration cycles for RTL tasks with clear metrics (defect reduction, time to lint clean, % auto‑fix suggestions accepted), and a stable MLOps path for continuous improvement.
How we work (Security & privacy by design) Customer and internal design data remain within private AWS VPCs; access via IAM roles and audited by CloudTrail; all artifacts encrypted with KMS.
No public internet calls for sensitive workloads; Bedrock access via VPC interface endpoints/PrivateLink with endpoint policies; SageMaker and/or EKS run in private subnets.
Data pipelines enforce minimization, tagging, retention windows, and reproducibility; DLP scanning and redaction are first‑class steps.
We produce model cards, data lineage, and evaluation artifacts for every release.
Tech you’ll touch Modeling: PyTorch, HF Transformers/PEFT/TRL, DeepSpeed/FSDP, vLLM, TensorRT LLM.
AWS & MLOps: Amazon Bedrock (Anthropic and other FMs, Guardrails, Knowledge Bases, Runtime APIs), SageMaker (Training/Inference/Pipelines), MLflow/Weights & Biases, ECR, EKS/KServe/Triton, Step Functions.
Platform/Security: S3 + KMS, IAM, VPC/PrivateLink (incl. Bedrock), CloudWatch/CloudTrail, Secrets Manager.
Tooling (nice to have): HDL toolchains for compile/simulate/lint, vector stores (pgvector/OpenSearch), GitHub/GitLab CI.
#J-18808-Ljbffr