LMTS - Lead AI/ML Engineer (LLM Post-Training)

Job description

To get the best candidate experience, please consider applying for a maximum of 3 roles within 12 months to ensure you are not duplicating efforts. Job Category Software Engineering Job Details About Salesforce Salesforce is the #1 AI CRM, where humans with agents drive customer success together. Here, ambition meets action. Tech meets trust. And innovation isn’t a buzzword — it’s a way of life. The world of work as we know it is changing and we're looking for Trailblazers who are passionate about bettering business and the world through AI, driving innovation, and keeping Salesforce's core values at the heart of it all. Ready to level-up your career at the company leading workforce transformation in the agentic era? You’re in the right place! Agentforce is the future of AI, and you are the future of Salesforce. What You'll Be Doing & Why It Matters Salesforce has immediate opportunities for engineers who want their work to shape how trusted enterprise AI is trained, evaluated, and delivered. You will work with a group of world-class engineers to train and post-train open-weight language models for demanding enterprise workflows. This is a model-training role, not a wrapper around a hosted LLM and not a prompt-engineering-only role. You will own the complete post-training loop: data generation, supervised fine-tuning, preference or reinforcement learning when appropriate, rigorous evaluation, and production serving. You will materially improve the behavior of models, not merely integrate model APIs. The role encompasses ML research engineering, data generation, fine-tuning, evaluation, inference systems, and technical leadership. You will own code, experiments, evaluation evidence, and production behavior. We value demonstrated engineering judgment and reproducible results over credentials, paper claims, or an extensive keyword list. We pride ourselves on building high-quality, maintainable systems that make the model more capable without compromising trust, safety, or operability. We embrace the hybrid model and celebrate the individual strengths of each team member while encouraging everyone to grow into the best version of themselves. We believe autonomous teams with the freedom to make decisions empower our people, products, company, and customers to thrive. Core Responsibilities & Scope Train, adapt, and serve capable, reliable open-weight language models for enterprise workflows. Own fine-tuning experiments from baseline through release: choose training objectives, run SFT and LoRA/QLoRA recipes, compare checkpoints against rigorous held-out evaluation, and diagnose failures before the next run. Generate high-quality domain training data at scale, including synthetic instruction data, multi-turn conversations, reasoning and tool-use trajectories, structured outputs, counterexamples, preference pairs, and difficult evaluation cases. Build the data engine behind post-training: source-grounded authoring, quality filtering, deduplication, contamination controls, dataset versioning, train/validation/test splits, packing, tokenization, and loss masking. Design and run post-training experiments beyond SFT when justified, including DPO, ORPO, GRPO, RLVR, distillation, and other preference or reinforcement-learning methods. Match the method to a measured capability problem and a defensible reward signal. Design and operate model-quality evaluation: held-out task suites, model-based and human review, tool-use and structured-output accuracy, multi-turn reliability, safety regressions, ablations, and error analysis. Investigate model behavior deeply: identify whether a failure comes from data quality, training objective, model capacity, decoding, context handling, tool interface, or serving behavior, then produce evidence for the corrective action. Build and ship high-quality, production-grade software using modern engineering practices, with AI as a core part of your development workflow. Build reliable model-serving and inference capabilities, balancing quality, latency, throughput, memory use, and operational reliability. Build robust integrations between models, tools, and enterprise systems, including structured outputs, function calling, and multi-step agent behavior. Make architecture and design decisions with strong attention to model quality, reproducibility, performance, security, scalability, and future extensibility. Improve observability, traceability, failure analysis, test coverage, and developer workflows across the training and serving lifecycle. Work with domain experts to turn real workflows, product behavior, APIs, and failure cases into independently verifiable training and evaluation material. Protect customer and company data: use approved sources only, preserve provenance, and keep credentials and sensitive data out of training and model prompts unless specifically authorized. Review model, data, and application changes for factual correctness, quality, security, performance, and operational risk. Mentor engineers and provide technical guidance that raises the quality bar across the team. Minimum Requirements 10+ years of software engineering experience, including technical leadership on production systems. Strong Python engineering experience building and maintaining reliable production systems. Hands-on ownership of LLM post-training. You have trained or materially improved open-weight language models through SFT and LoRA/QLoRA. Experience with preference optimization, RL, distillation, or continued pre-training is strongly valued. API integration alone does not qualify. Hands-on ownership of training-data generation and curation. You have produced, filtered, reviewed, and versioned task-specific datasets, including synthetic instruction data, conversations, reasoning traces, tool calls, preference data, or structured outputs. Strong practical understanding of the post-training stack: chat templates, tokenizer behavior, sequence packing, loss masking, optimizer and learning-rate schedules, checkpointing, distributed or memory-efficient training, and inference tradeoffs. Experience creating and evaluating reasoning, tool-use, or structured-output training data. You understand the difference between a plausible trace and a correct, useful outcome, and know how low-quality traces degrade a model. Experience with model evaluation: held-out datasets, error taxonomy, task and quality metrics, experiment design, ablation studies, regression analysis, and human-review workflows. Strong software-engineering fundamentals: API design, testing, debugging, observability, version control, and clear technical writing. Ability to investigate ambiguous failures across model behavior, data, prompts, retrieval, tools, and backend services. Demonstrated leadership in architecture, design reviews, code reviews, and mentoring engineers. Comfortable working across application and ML layers instead of staying within a narrow specialty. A demonstrated, genuine AI-first approach to engineering — using AI to move faster, build fluency across the stack, and contribute well beyond your core specialty. Experience using AI tools (e.g., Claude Code, GitHub Copilot, Codex, Cursor, etc.) in development workflows. Advanced prompt engineering skills and the ability to write precise, structured prompts and cultivate the system context that makes AI outputs reliable, secure, and production-ready. Preferred Qualifications Experience building integrations with enterprise platforms, APIs, data systems, identity controls, or tool-using AI applications. Experience with complex enterprise workflows, such as CRM, commerce, finance, service, supply chain, or regulated-industry systems. Deep experience with preference optimization or reinforcement-learning methods for LLMs, such as DPO, ORPO, GRPO, PPO, or verifiable-reward training. Candidates must be able to explain the reward/preference data, training stability, and failure modes, not just name a framework. Res