Fine-Tuning Specialist

A meticulous model sculptor who transforms base foundation models into purpose-built inference engines that outperform their general-purpose ancestors on domain-specific tasks — with the deep learning rigor to know when fine-tuning is the right lever versus prompting, RAG, or distillation.

- The DISTILL Fine-Tuning Methodology — 7-stage framework from bottleneck diagnosis to lifecycle management

- Data curation playbooks: deduplication, toxicity filtering, synthetic generation, mixture ratio optimization

- PEFT strategy (LoRA, QLoRA, adapters) with rank selection grounded in task complexity analysis

- Evaluation suite design that mirrors production conditions and detects overfitting, memorization, regression

- Hyperparameter tuning strategies: learning rate scheduling, warmup, per-layer differential rates, multi-task weighting

- Alignment and safety-aware tuning across RLHF, DPO, Constitutional AI, red-team integration

- Production deployment frameworks: quantization, serving integration, A/B testing, rollback procedures

- Technology stack guidance: Hugging Face Transformers, Axolotl, DeepSpeed, vLLM, LangSmith, Weights & Biases

- Diagnosis framework preventing wasted compute by separating fine-tuning from prompting and retrieval needs

Drop into Claude, ChatGPT, Cursor, or any AI tool. Bring your real fine-tuning problem — a performance gap you need to close, a data curation challenge, a PEFT versus full-tuning trade-off, a safety regression concern. It thinks like an ML engineer who's survived catastrophic forgetting, reward hacking, and overfitting in production.

Bundles or prompts related to machine learning operations and model development strategy.