$ whoami
saisha_shetty
> PhD student @ UC Davis · trustworthy AI for healthcare
>
type help or a section name · press enter for console · or scroll ↓
$ cat about.md
## whoami
I build clinical AI that knows when it doesn't know. I'm a PhD student in Computer Science at UC Davis, working on trustworthy AI for healthcare — LLM evaluation, uncertainty calibration, and synthetic data.
I do this research with Prof. Vladimir Filkov in the AI for Health Initiative (AI4H), where I lead RadAnnotate — an LLM-powered framework for reliable radiology report annotation, combining entity-specific models, synthetic data, and confidence-based selective automation.
I care less about topping a benchmark and more about a model not confidently lying to a doctor at 2 a.m.
// everything else is a command — try publications, projects, experience, news, or cv.
$ contact --all
$ git log --recent # latest news
2026 · 12★First-author poster accepted at RSNA 2026, Chicago. — Structured Multi-Agent Clinical Reasoning with Automatic Task Decomposition.
2026★RadAnnotate accepted to the AMIA 2026 Amplify Informatics Conference.
2026 · 10★First-author poster accepted at CAIMI26 (SIIM · UPenn). — Multi-Agent Reasoning Enhances Local Small-Model Performance in Medicine.
2026 · 10+Reviewed for the AMIA 2027 Amplify Informatics Conference.
2026 · 09+Reviewed for ML4H 2026 (AHLI symposium).
2025 · 07+Joined ELTA AI as an AI Engineer.
2025 · 03★Honored by UC Davis as an Exceptional Woman in Computer Science.
2024 · 04+Started as a Graduate Student Researcher with Prof. Filkov / AI4H @ UC Davis.
// this feed is hand-committed. more history lives in the CV.
$ ls -l publications/
papers/
MARC v1: An Open-Source Multi-Agent Framework for Clinical AI Reasoning and Coordination
[paper][code]
An open-source multi-agent framework that coordinates specialized agents to decompose and reason through clinical tasks, lifting small, locally-run models in medicine.
Reliable annotation of radiology reports using entity-specific models, RAG-based synthetic data, and confidence-based selective automation. F1 0.89, with ~67% of reports automated at expert-level precision.
A fine-tuned U-Net for segmenting polyps in colonoscopy imagery. 96.8% accuracy and Dice 0.76.
Automated Identification of Make and Model of Total Wrist Replacement Implants using Deep Learning
[paper]
A transfer-learning pipeline (ResNet50V2) that identifies wrist implant make/model from radiographs to support pre-operative planning. 95.1% accuracy, AUC 0.9959.
posters & abstracts/
Structured Multi-Agent Clinical Reasoning with Automatic Task Decomposition and Pipeline Generation
[code][poster]
A multi-agent system that decomposes a clinical question into sub-tasks and generates its own reasoning pipeline.
Multi-agent reasoning that lifts small, locally-run models on medical tasks. Abstract to appear in the JIIM supplement (Springer Nature).
$ ./projects --featured health
health/
MARC-v1multi-agent clinical reasoning
RadAnnotateclinical LLM · radiology
Auto-annotates radiology reports with entity models + calibrated confidence — F1 0.89, 67% safely automated — so experts only review what actually needs a human.
SegRetinomedical imaging · segmentation
U-Net for retinal image segmentation with pixel-level accuracy to support faster diagnosis, using augmentation + post-processing to generalize across scans.
other/
$ git log --oneline --experience
●Graduate Student Researcher — UC Davis · AI4HApr 2024 — now
Lead RadAnnotate: RAG-enhanced synthetic data + multi-model consensus scoring for radiology reports — F1 0.89 while safely automating 67% of cases. Calibrate model confidence with isotonic regression (Brier ↓ up to 0.13).
●NLP / LLM Data Scientist Intern — Dandelion HealthJul — Sep 2026
Built an end-to-end NLP pipeline to extract inflammatory bowel disease complications from ~150K radiology reports — SQL-based candidate retrieval, LLM structured extraction, and human-in-the-loop validation.
●AI Engineer — ELTA AIJul — Sep 2025
Built an LLM-driven travel-risk system (Google MCP + DynamoDB) giving real-time 0–100 safety scores at 95% accuracy, plus a Guardrails + regex sanitization layer that cut unsafe text by 80% in production.
●Machine Learning Engineer — Sentry AI2021 · 2023
Improved small-object detection by 40% with slice-aided inference for real-time surveillance; built ResNet152 + Detectron2 pipelines for vehicle & license-plate recognition and lifted false-alert precision by 25%.
●Head Teaching Assistant — UC Davis · ECS 154A2023 — 2024
Led the TA team for Computer Architecture — ran discussion sections, office hours, and grading (also TA'd ECS 36B, C++ / OOP).
// earlier roles — MLPal, UVic Energy-in-Cities, SRM Biotech — live in the CV.
$ open cv.pdf
The complete record — education, publications, awards, service. ↳ download cv.pdf