jatin avinash salve

Research Assistant — AI Systems, LLM Evaluation & Infrastructure

University of Florida College of Medicine

Building reproducible clinical LLM evaluation infrastructure on UF HiPerGator.

• Built LLM experimentation infrastructure with Slurm, Apptainer, Python, Neo4j, and Graphiti for reproducible runs and persistent graph memory.
• Benchmarked direct, long-context, vanilla RAG, time-aware RAG, hierarchical, rolling-memory, and graph temporal-memory summarization.
• Built a six-stage verification pipeline for temporal Graphiti and Neo4j memory, evaluated against 1,682 hand-audited atomic facts.
• Contributed to a permission-aware React, TypeScript, and FastAPI application with retrieval, validation, structured logs, Docker, ECS, and CloudWatch.

Machine Learning Engineer — Production ML Systems

ICICI Bank Pvt Limited

Built and operated production inference and ranking systems serving more than one million daily requests.

• Built FastAPI, Docker, Kubernetes, and ECS inference services; improved p95 latency by 5% through feature retrieval and serving optimizations.
• Developed feature-store ranking and propensity workflows that increased click-through rate by 9%.
• Owned preprocessing, retraining, deployment, rollback, CI/CD, integration testing, telemetry, and production debugging.
• Built Python A/B and performance analyses across latency, throughput, relevance, conversion, and customer cohorts.

Research Assistant — Multimodal Foundation Models

IIT Patna AI-ML-NLP Lab

Developed and evaluated multimodal encoder-decoder systems for clinical document summarization.

• Extended BART with image fusion and decoder cross-attention.
• Benchmarked three datasets against BART, GPT-4V, and multimodal baselines using ROUGE, BLEU, METEOR, statistics, and expert review.
• Profiled ResNet, VGG, ViT, and fusion choices; documented failures and ablations for an ACL 2024 long paper.

Software Developer & Open-Source Mentor

Google Summer of Code — Joomla CMS

Developed the Guided Tour feature in Joomla 4 core for a platform used by 2M+ users.

• Worked across MVC architecture, plugins, testing, code review, and documentation, then returned as an open-source mentor.

Latency-Aware Speculative Decoding & Draft-Model Distillation

Built a correct speculative decoding engine with drafting, verification, rejection correction, bonus tokens, and KV caching; verified exact greedy equivalence.

• Reduced 64 target passes to 22 verification passes and reached 2.91 tokens per target pass.
• Profiled 0.79x–1.09x performance versus greedy decoding and found the draft model consumed 62% of GPU time.
• Improved cached greedy throughput from 92.96 to 102.06 tokens/s; distillation improved acceptance by 5% on GSM8K and 4% on CNN/DailyMail.

Verifiable Preference Optimization for Reasoning LLMs

Generated 1,800 verified reasoning traces and 330 preference pairs for LoRA DPO with Hugging Face TRL.

• Established a 61–63 generated tokens/s baseline on NVIDIA L4.
• Diagnosed truncation and answer-parsing label noise and redesigned validation for reliable evaluation.