Software Engineer building production AI systems

Jatin Kulkarni

I build AI systems at the boundary between models and real-world software. My current work focuses on healthcare voice agents, multi-agent orchestration, evaluation, latency, and reliable routing around probabilistic model behavior - building on experience in SageMaker ML infrastructure, full-stack products, and multimodal research.

AI Agents & Voice LLM Evaluation Distributed AI Systems ML Infrastructure Seattle, WA
View Resume Connect on LinkedIn

Selected Work

Reliable AI systems, from model behavior to production infrastructure

AI Agents

Production Healthcare Voice AI

Built multi-agent orchestration for latency-sensitive healthcare conversations, combining model-driven behavior with persisted state, explicit routing controls, and evaluation.

  • Reduced first-turn latency from approximately 14s to 5s in test flows
  • Added LLM-simulated-patient evaluations and regression gates
  • Implemented parallel execution and deterministic routing fallbacks
Read the case study

Research

HEDWIG: Geospatial Vision-Language Research

Built a ViCLIP-based geolocation system that learns richer geospatial embeddings from multi-frame panoramic imagery and captions.

  • Reduced median top-1 prediction error by over 1,600 km vs. CLIP baseline
  • Increased predictions within 750 km by nearly 4x
  • Studied embedding strategies and geographic-resolution tradeoffs
View the research project

Full-Stack AI

FinQ-RAG: Financial Document Question Answering

Built a full-stack application for grounded question answering over financial PDFs, from retrieval and model integration to the user-facing product.

  • FastAPI service with SentenceTransformers and FAISS retrieval
  • Hugging Face generation grounded in document context
  • Angular and Tailwind interface for end-to-end use
See the project summary

About

Turning model capability into dependable product behavior.

My path into production AI started with a computer science and machine learning foundation, then moved through multimodal research, early LLM and RAG work, and production ML infrastructure at AWS. Today I work on healthcare voice-agent systems where model behavior has to coexist with latency constraints, application state, routing rules, evaluation, and real user needs.

The part of AI engineering I find most interesting is everything required to make a capable model dependable: evaluation, state, routing, observability, safeguards, and user experience. My background in business, entrepreneurship, and design keeps that engineering grounded in whether the system actually works for the person using it.

View My Resume

What I Work On

Reliable AI systems across agents, evaluation, and infrastructure

Multi-agent orchestration, deterministic controls around probabilistic behavior, conversational latency, persisted state, observability, and full-stack product integration.

How I Think

Technical correctness and user experience are the same system

I like finding the critical invariant, measuring the real failure mode, and designing the smallest dependable path through model, backend, and product constraints.

Recent Signals

  • Reduced first-turn voice-agent latency from approximately 14s to 5s in test flows
  • Built LLM-simulated-patient evaluations and regression gates for multi-agent workflows
  • Added deterministic recovery paths around probabilistic routing decisions

Experience

Building across AI products, infrastructure, and research.

My work has moved from multimodal research and applied LLM systems to production ML infrastructure and, now, reliable agent systems for latency-sensitive conversations.

Healthcare AI & SageMaker

Amazon Web Services / Amazon

Software Development Engineer, AI Platforms & Healthcare AI

Seattle, WA

July 2025 - Present

Current focus

Healthcare AI / AI Agent Systems

  • Built multi-agent orchestration for healthcare voice AI using Python, Java, Amazon Bedrock, LangGraph/ReAct, and persisted session state across patient verification, appointment, and medication workflows.
  • Developed orchestrator evaluations with LLM-simulated patient personas and LLM-as-judge scoring for task completion and conversational quality, adding regression thresholds to existing nightly pipelines.
  • Reduced first-turn voice-agent latency from approximately 14s to 5s in test flows by prewarming sessions in parallel and overlapping agent preparation with routing work.
  • Implemented parallel primary/background agent execution and deterministic fallbacks around incomplete or inconsistent probabilistic routing decisions.
  • Built tooling to correlate multi-service logs and isolate latency across agent execution, orchestration, forwarding, and response delivery.

Earlier chapter

SageMaker AI Platforms / ML Infrastructure

  • Extended backend APIs, validation, and capacity-routing flows for SageMaker Training Plans across training and inference workloads.
  • Reduced configuration and deployment turnaround from 6-10 days to 1-2 days and enabled zero-touch region expansion through dynamic configuration.
  • Analyzed 290+ support and limit-increase cases and helped redesign capacity limits, reducing customer friction by about 96%.
  • Built React/TypeScript functionality in SageMaker Studio to surface capacity, remaining time, instance type, and cost information.
Public SageMaker product reference

Visual Computing Group

Cornell University

Researcher

New York, NY

Fall 2024 - Spring 2025

  • Researched image geolocation using hierarchical geocells and ViCLIP-based vision-language representations.
  • Ran ablation studies across embedding strategies, training setups, and geographic resolutions.
  • Built an experimentation workflow for multimodal representation learning and large-scale retrieval.

AI Research & Development

Aristocrat Technologies

Intern

Austin, TX

January 2024 - June 2024

  • Built LLM-based RAG systems with LlamaIndex and LangChain for test-case generation, failure diagnosis, and root-cause analysis.
  • Tested retrieval configurations and prompt strategies to improve generated test quality and reduce hallucinations.
  • Developed defect-prediction models from historical testing data to improve QA prioritization and reduce flaky-test impact.

Technical

Kellogg Brown & Root International, Inc.

Intern

Houston, TX

May 2023 - August 2023

  • Implemented an automation pipeline for migrating enterprise documents into Azure Cognitive Search.
  • Configured AWS IAM and Azure Active Directory access patterns for new projects and roles.
  • Analyzed more than 600,000 support tickets to identify trends and improve IT department efficiency.

Software Engineering

Responsible Artificial Intelligence Institute

Intern

Austin, TX

January 2022 - December 2022

  • Helped redesign and scale the React/Firebase architecture behind RAI Collab.
  • Built 7 of 17 responsible-AI tools for the platform.
  • Built the AI Regulatory Tracker to track 170+ global AI regulations.

Undergraduate Research

UT Austin Autonomous Robotics

Researcher

Austin, TX

August 2021 - December 2021

  • Studied perceived safety in human-robot interaction experiments involving Boston Dynamics Spot.
  • Used ROS, RVIZ, Azure Kinect, and SLAM workflows to model test spaces and track subjects in real time.
  • Combined technical implementation with experimental analysis to understand comfort and safety tradeoffs.

Projects

AI systems across agents, product engineering, and research.

Case studies and projects spanning production healthcare AI, multimodal learning, retrieval-augmented generation, ML infrastructure, and full-stack product work.

Current Work

Production Healthcare Voice AI

Multi-Agent Orchestration, Evaluation, Reliability, and Latency

Built production orchestration around healthcare voice agents, combining probabilistic model behavior with deterministic state, routing controls, evaluation, and observability.

  • Reduced first-turn latency from approximately 14s to 5s in test flows
  • Developed LLM-simulated-patient evaluations and regression gates
  • Implemented parallel execution and deterministic routing recovery paths
View full project page

Fall 2024 - Spring 2025

HEDWIG: Learning Geospatial Embeddings for Large-Scale Retrieval

Representation Learning, Multimodal Retrieval, and Geolocation

Built a ViCLIP-based geolocation system that learns richer geospatial embeddings from multi-frame panoramic imagery and captions.

  • Reduced median top-1 prediction error by over 1,600 km vs CLIP
  • Increased predictions within 750 km by nearly 4x
  • Improved top-1 retrieval quality across distance thresholds
View full project page Read full paper

Spring 2024

Multimodal Medical Image Classification using CLIP and ResNet

Multimodal Learning and Interpretability

Explored multimodal and image-only approaches for diabetic retinopathy classification, comparing semantic alignment against fine-grained visual discrimination.

  • ResNet-50 achieved 92.52% accuracy
  • Two-stage CLIP approach achieved 89.31% accuracy
  • Used Grad-CAM and t-SNE to analyze explainability and embedding structure
View full project page Read full paper

2025 - Present

AWS AI Platforms: Scalable ML Infrastructure and Production Systems

Production ML Infrastructure and Reliability

Software engineering work on SageMaker Training Plans infrastructure for reserved-capacity procurement, allocation, and validation across ML training and inference workloads.

  • Reduced deployment/configuration turnaround from 6–10 days to 1–2 days
  • Reduced customer friction by ~96% through data-driven reserved-capacity limit redesign
  • Contributed to inference-related Training Plans workflows for reserved GPU/accelerator capacity
View full project page Related AWS blog post

Fall 2024

Handwritten Equation Recognition using CNNs, Vision Transformers, and Seq2Seq Models

Computer Vision + Sequence Modeling

Built a multi-stage pipeline to convert handwritten equations into LaTeX using visual recognition and sequence generation models.

  • ViT-Base reached 72.87% symbol recognition accuracy
  • Seq2Seq model achieved 66.19% exact match
  • BLEU score of 0.8428
View full project page Read full paper

Spring 2025

FinQ-RAG

Full-Stack Retrieval-Augmented Generation System

Built a full-stack application for answering natural language questions over financial PDFs using FastAPI, Angular, FAISS, and Hugging Face models.

  • Implemented context-grounded retrieval with SentenceTransformers and FAISS
  • Built the backend API with Python and FastAPI
  • Delivered the user-facing workflow with Angular and Tailwind CSS

Spring 2024

Neural Network Final Project: Harmonizing Genres

CS 342 Neural Networks

Predicted music genres by combining CNN-based spectrogram understanding with RNN/LSTM sequence modeling for richer audio representation.

View paper

Contact

Let's compare notes on building dependable AI.

I enjoy conversations about agent systems, evaluation, applied ML, infrastructure, and the product decisions that turn capable models into useful software.

LocationSeattle, WA

Emailj10.kulkarni@gmail.com

LinkedInlinkedin.com/in/jatin-kulkarni

GitHubgithub.com/jatinkulkarni