Rohan Sitaniya

Agentic workflow automation for enterprise and citizen-developer customers · Founding AI team (1 of 3), reporting to the CTO · Remote • iPaaS AI Assistant: Owned the multi-agent system behind Jitterbit's natural-language integration builder through production rollout: a LangGraph Planner retrieves from 400+ Marketplace recipes and hands a Pydantic-validated plan to an Executor with checkpointed, resumable execution. • Retrieval & context: Each agent got its own context and model, and a token-aware memory layer summarizing between turns kept long sessions bounded, cutting LLM spend ~40% with answer quality held. Pinecone-backed retrieval improved context precision 60%+ and cut latency 30%. • APIM AI Assistant: Shipped Jitterbit's natural-language API lifecycle agent, handling 10K+ daily interactions so users create and publish APIs without filing a ticket with engineering. A provider-agnostic adapter spans OpenAI, Bedrock and open-source models. • Operations: DeepEval response checks, hallucination and bias detection, Langfuse request-level tracing and model-drift monitoring, on CI/CD-driven Kubernetes microservices so the controls shipped with every release.

Jitterbit

AI Labs · Emerging AI Products · Model Risk Management • GenAI Governance: Scoped and led a 5-member team building a GenAI evaluation framework for production reviews: 20+ quantitative and 30+ qualitative checks covering hallucination, bias, toxicity, PII, faithfulness and auditability, run against representative production data and wired into automated model gating. Cut gen-ai governance time by ~60%. • Fraud Embeddings: Designed Transformer-based embedding architectures that outperformed the legacy RNN representations by ~40bps, and built GPT-based merchant embeddings that went into production fraud pipeline. • AI Underwriting Automation: The commercial new-accounts team was reading bank statements manually. I scoped the problem with them, made the case for an NLP approach over adding reviewers, and shipped transaction-categorization models at ~85% accuracy to score the credit worthiness of the applicant, taking that review cycle from days to hours. • Risk Intelligence: Built an LLM-powered pipeline for the credit-risk unit, reading real-time news and external signals into automated early-warning alerts preventing sudden, unforeseen multi-million dollar delinquencies.

American Express

• Fine-tuned BERT and GPT-2 for personalized query auto-completion over 1.5 million business analytics queries.

Algonomy

• Genomic analysis of Axolotl RNA-Seq data to identify regenerative epigenomic markers, mapped to the human genome.

National Taiwan University