Headline

AI Engineer | Building Production LLM Systems & Agentic AI

About

I build LLM systems that have to survive contact with a bank. At State Street I lead work on the Emerging Tech team: agentic workflow platforms, RAG pipelines over fund prospectuses, and a Text-to-SQL copilot that lets people ask financial data questions in plain English. The interesting part is rarely the model. It's the guardrails, observability, latency and evaluation, the unglamorous scaffolding that decides whether a demo turns into something a regulated business will actually run. Outside work I pretrain small language models from scratch and build tooling for evaluating agents, mostly to understand the stack from the bottom up instead of just the API surface. Co-author on two papers applying Graph-of-Thoughts reasoning to Text-to-SQL (FLLM 2025, CSCI 2025). Portfolio and selected work: himanshunakrani.me

Location

Hyderabad

AI stack

opencode

claudecode

Work experience

Emerging lead at State Street, Apr 2026 - Present · 5 mos, On-site, Leading agentic AI platform development on State Street's Emerging Tech team. • Spearheaded "Agent Forge", an enterprise no-code agentic workflow platform managing the complete agent lifecycle: visual design, deployment, monitoring and evaluation of multi-agent systems, with centralized agent registration, discovery and role-based access control (RBAC) for governance. Adopted by 5+ business teams running 50+ agent workflows in production. • Designed and implemented OTEL-compliant model observability, AI guardrails with input/output validation and content filtering, and automated evaluation pipelines for Responsible AI and regulatory compliance across all registered agents.

Senior Associate at State Street, Apr 2024 - Apr 2026 · 2 yrs 1 mo, On-site, Building production LLM applications over enterprise financial data. • Architected and deployed "Alpha Copilot", an enterprise LLM-powered Text-to-SQL chatbot enabling natural language queries over structured financial data for 100+ internal users, with LLM-generated visualizations, interactive charts and summaries to accelerate decision-making. • Built a production RAG system for extracting actionable insights from fund prospectuses, engineering a document pipeline with intelligent chunking, semantic embeddings and vector database integration that achieves sub-second retrieval. • Fine-tuned GPT-4o mini and GPT-4.1 mini (Azure OpenAI) and Llama 3.1-8B (Databricks) using supervised fine-tuning and PEFT/LoRA, owning the end-to-end LLMOps pipeline from data augmentation through training and inference optimization. • Engineered scalable AI inference infrastructure with FastAPI and SQLAlchemy; caching and latency optimization reduced average response time by 75%, from 25-30s to 6-8s. • Researched schema-linking and few-shot prompting strategies for Text-to-SQL generation, improving SQL accuracy by 25%. This work was published at IEEE FLLM 2025 and CSCI 2025.

Associate 2 at State Street, Jul 2023 - Apr 2024 · 10 mos, On-site, Backend and platform engineering for AI-driven wealth products. • Led backend development for the WealthAI application, building a real-time portfolio narrative generation system that processes optimizer outputs and delivers personalized investment insights at scale, working with cross-functional Agile teams. • Architected automated CI/CD pipelines with Harness and Docker and drove unit test coverage above 95%, streamlining deployment workflows and reducing production incidents. • Built and maintained Python and FastAPI services underpinning downstream AI applications.

Machine Learning Summer School at Amazon, Jul 2022 - Jul 2022 · 1 mo, Remote

Embedded System & IoT Intern at Green Field Control System (I) Pvt Ltd, Jun 2022 - Jul 2022 · 2 mos, Gandhinagar, Gujarat, India

Machine Learning Developer at Technocolabs Softwares, Jun 2022 - Jul 2022 · 2 mos, Remote

Education

Bachelor of Technology - BTech, Electronics and Communications Engineering, Nirma University

Public links

github: https://github.com/himanshu-nakrani