Pavel Tishchenko

Trained and fine-tuned Transformer models using PyTorch and Hugging Face. Implemented distributed training with DDP, FSDP and DeepSpeed (ZeRO-2/3). Improved macro-F1 from 0.68 to 0.84 and rare-class F1 by 18%. Built reproducible training and evaluation pipelines.

Fine-tuned open-source LLMs on domain-specific datasets using PyTorch and Hugging Face. Optimized production inference with vLLM, batching and quantization, reducing p95 latency from 2.3s to 1.1s and increasing throughput from 18 to 42 requests/sec. Automated offline regression evaluation and CI/CD quality checks.

Developed DPO fine-tuning pipelines using Hugging Face TRL, improving LLM win rate from 51% to 68%. Built multi-step AI agent workflows with LangGraph, increasing task success rate from 73% to 89%. Implemented structured tool calling, error handling and evaluation pipelines for LLM-based banking automation.

Contributed to candidate retrieval work within a recommender systems team. Trained two-tower retrieval models, performed offline evaluation, and prepared experimental solutions for product validation. Improved Recall@100 and NDCG@10 versus baseline, and built a reproducible offline evaluation pipeline with metrics and segmented slices. Performed error analysis to identify cold start, popularity bias, and low candidate diversity, informing follow-up experiments and A/B preparation.