•Own the Model Adaptation Lifecycle: Drive end-to-end adaptation of SLMs to specialized domain tasks, choosing the right base models, training strategies, and architectures.
•Apply Efficient Fine-Tuning: Implement parameter-efficient fine-tuning (PEFT) methods with a mathematically grounded understanding of low-rank adaptation (LoRA, QLoRA, rank selection, parameterization, scaling, and accuracy–efficiency trade-offs).
•Select Strategy per Task: Determine and sequence optimal adaptation strategies - including Continued Pre-Training (CPT), Supervised Fine-Tuning (SFT), and RL/preference tuning (DPO, GRPO, RLHF).
•Curate & Synthesize Data: Clean, curate, format, and generate high-quality datasets, including synthetic data generation when domain data is scarce.
•Design Functional Evaluations: Build task-specific benchmarks grounded in functional, executable correctness rather than loss curves or perplexity alone, maintaining regression tracking across versions.
•Leverage Knowledge Transfer: Use larger models as teachers to transfer capabilities to SLMs via distillation, structured pruning, and quantization.
•Deploy & Optimize Serving: Stand up and optimize inference infrastructure across GPU, CPU, and edge targets using frameworks such as vLLM, TGI, Ollama, llama.cpp, or TensorRT-LLM.
•Pipeline Reproducibility & Research Execution: Break target tasks into milestones with clear criteria, maintaining reproducible pipelines with data/config versioning, checkpointing, and experiment tracking.
•Translate recent literature into reproducible experiments.
