Onkar Salvi
Senior Research Engineer · AI Agents / LLMs / ML / NLP
Research Engineer with 7+ years building, training, and scaling ML/NLP systems hand-in-hand with research scientists. Depth in LLM finetuning (LoRA/PEFT), contrastive & representation learning, agentic systems, and model optimization — turning research ideas into production systems that process 600M+ messages daily. Published in AAAI; 1st place at the DSTC7 dialogue challenge.
About Me
I'm a Research Engineer with deep expertise in AI agents, Large Language Models, and Natural Language Processing. Currently a Senior Research Engineer at Dataminr, I work hand-in-hand with research scientists to turn ideas into production systems — from multi-agent workflows and LLM finetuning to model optimization and serving infrastructure.
My work spans the entire AI lifecycle: LoRA/PEFT finetuning, contrastive & representation learning, agentic (semantic + keyword) retrieval, and cost-efficient inference at scale. I have a strong bias toward clean, reusable research infrastructure and fast, ablation-driven experimentation.
I've published in AAAI, placed 1st at the DSTC7 dialogue challenge, and served as a reviewer for the AIWILD Workshop at ICML and ICLR 2025.
Experience
- Architected multi-agent workflows for the Knowledge Accumulation pipeline powering Tailored Intelligence, Dataminr's flagship offering — agents autonomously gather client-specific context to enable per-client impact analysis of every alert
- Re-architected app search from keyword lookup to agentic hybrid (semantic + keyword) retrieval — a Query Analyzer for intent/multi-hop decomposition over an OpenSearch vector store scaled to 100M+ vectors
- LoRA fine-tuned LLaMA-3.1-8B in a two-stage architecture (contrastive-learning pre-filter + LLM deep analysis) for real-time threat detection across 600M+ daily messages; ablations lifted F1 by 23% (65% → 88%) at sub-second latency
- Built an agent-assisted annotation pipeline supplying cultural context for threat classification, cutting annotation cost ~66% vs. human-only labeling
- Shipped a Sales Enablement app combining a web-search agent with a RAG workflow — 3× email open rate and 60% lift in response rate
- Built the company's first Model Context Protocol (MCP) server (web search, document parsing, internal API access); integrated Arize Phoenix and Langfuse for LLM/agent observability
- Built and maintained the in-house model-serving framework (gRPC + REST proxy) deploying NLP/CV/LLM models on Kubernetes (AWS EKS), with metric-aware autoscaling on queue length, latency, and request volume
- Led migration of 40+ AI services from ECS to Kubernetes and cut inference cost 70% across 14+ models via AWS Inferentia, graph optimization, and dynamic batching — $800K annual savings
- Researched and developed ML/DL models for natural language understanding, chatbots, machine reading comprehension, and NLG
- 1st place, DSTC7 Dialog System Technology Challenge (knowledge-grounded conversation) — work published at AAAI
- Built a zero-shot classification model reaching 98% accuracy on seen and 90% on unseen classes, eliminating retraining on new data
- Applied BERT + K-Means clustering over large call-transcript corpora to group semantically similar utterances; used Celery for asynchronous model retraining
- Owned a ranking model that ordered related news drawn from multiple sources
Skills & Technologies
🤖 Deep Learning & Research
⚡ Model Optimization & Serving
☁️ Infrastructure & MLOps
💻 Languages
Education
Publications & Service
- J. Zheng, O. Salvi, J. Chan. "Candidate Attended Dialogue State Tracking using BERT." AAAI, 2020.
- J. Zheng, S. Kasturi, M. Lin, X. Chen, O. Salvi, H. J. Wang. "OneConn-MemNN System for Knowledge-Grounded Conversation Modeling." AAAI, 2019.
- Reviewer, AIWILD Workshop, ICML 2025
- Reviewer, AIWILD Workshop, ICLR 2025