Home /
Senior Architect, Model Performance and Post-Training
– jobstoday.jobdekhi
Senior Architect, Model Performance and Post-Training
Other Jobs To Apply
No other job posts for this day.
Job Description:
Develop meticulous frameworks to measure and analyze coding model performance across code generation, debugging, code completion, repository-level reasoning, and multi-step software engineering tasks
Develop and optimize post-training pipelines for coding models, including supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), and outcome-based or process-based reward modeling
Analyze benchmark results, training dynamics, failure modes, and model behaviors to find opportunities for improvement
Improve model performance on sophisticated coding tasks through targeted training interventions, reward design, curriculum development, and iterative error analysis
Build scalable, reproducible infrastructure for training experiments, evaluation pipelines, and model performance analysis
Collaborate with research scientists, infrastructure engineers, and product teams to turn model insights into production-ready capabilities
Requirements:
Bachelor's degree (or equivalent experience) in Computer Science or AI
12+ years of meaningful work experience
Excellent programming skills in Python
Strong understanding of large language models, transformer architectures, and training and evaluation of generative models
Foundational understanding of reinforcement learning algorithms and implementation, including MDPs, policies, and reward design
Deep understanding of software engineering fundamentals, code generation, automated testing, and program correctness
Proven track record to lead model development life cycles, including dataset versioning, experiment tracking, and evaluation pipelines
Experience training or post-training state-of-the-art coding models or models designed for software engineering agents
Experience with execution-based rewards, verifiable rewards, reward hacking mitigation, and automated code evaluation
Familiarity with coding benchmarks and evaluation environments such as SWE-bench