ML systems
I build retrieval, inference, and model-lifecycle workflows that make ML useful beyond a notebook.
FastAPI · vector search · model serving
About Kiran Shrestha
I build reliable AI systems—from research to real-world use.
I'm an ML engineer and researcher working across AI products, retrieval and LLM systems, computer vision, and production ML infrastructure. I build the models, evaluation workflows, and software around them so technical ideas can become dependable tools.
Perspective
How I work
I like starting with a real problem, not a model. Sometimes that means helping someone understand a fraud score; other times it means making a research tool easier to use, finding a defect, or replacing a manual workflow. I first get clear on what needs to improve, then build and test the system around that.
Research taught me not to take a promising result at face value. I care about how a system is evaluated, where it breaks, and whether it holds up outside a notebook. That mindset carries into the retrieval pipelines, APIs, and interfaces I build.
Trajectory
Education, research, and engineering
2016–2020
B.S. Mathematics
Southeastern Louisiana University · Minor in Computer Science
2021–2025
M.S. Computer Science
Tulane University
2022–2025
Graduate Research Assistant
Three years of computer vision research at Tulane University
2023
Annotation-free X-ray defect localization
Completed FFT + SAM crack-detection research
2024–2025
Active computer vision research
Few-shot anomaly detection and interpretable image recognition
2025–2026
Applied ML systems
Built research, vision, fraud-explanation, and workflow products
Capability areas
Technical focus
Computer vision
My research focuses on interpretable recognition, few-shot anomaly detection, and annotation-free defect localization.
DINO · prototype learning · SAM
Research & evaluation
I frame research questions, compare methods, and evaluate model behavior with benchmark and validation workflows.
pAUROC · W&B · validation pipelines
Production engineering
I carry systems through APIs, containerization, deployment, and automation so they can be tested in practice.
Docker · AWS · CI/CD
Currently exploring
Interpretable and few-shot vision systems
Active research on making image-classification decisions easier to inspect and detecting industrial anomalies from limited normal examples.
Let's build something useful.
Explore the work, review my background, or start a conversation.