I'm a researcher and engineer at the University of Florida β been here long enough to know the hospital WiFi passwords. I split my time between two labs: IC3 (Intelligent Clinical Care Center) and PRISMAp (Precision & Intelligent Systems in Medicine).
The short version: I build the data infrastructure and systems that make clinical AI actually work β ETL pipelines, EHR data wrangling, embedding models into production, making sure the data is clean and the systems stay up. I've been part of 25+ papers across Nature Scientific Reports, JAMA Surgery, JMIR, and ICLR workshops. 300+ citations. Multiple NIH grants (R01, R21, K01, OT2) funded that work.
The thing is β someone has to take these models off the researchers' laptops and get them running in a real hospital. That's me. Models don't save patients. Systems do. So I build the plumbing, the guardrails, and the infrastructure that makes clinical AI deployable.
MS in Electrical & Computer Engineering, UF. Data Scientist III, Department of Medicine.
Our lab has a thing for fruit names. Here's what I helped build the data and systems for:
| Paper | TL;DR | Scale |
|---|---|---|
| MySurgeryRisk | "Will this surgery go wrong?" β multicenter validation | 508k surgeries, AUROC 0.95 |
| DeLLiriuM | LLM that predicts delirium from structured EHR | 104k patients, 195 hospitals |
| MANDARIN | Mixture-of-Experts for delirium + coma | 92k ICU patients |
| MANGO | Multimodal transformer for ICU acuity | Vitals + labs + notes fused |
| APRICOT-Mamba | State-space model for real-time ICU prediction | 75k patients, 147 hospitals |
| Temporal Cross-Attn | Dynamic EHR tokenization | ICLR 2024 Workshop |
| Federated Learning | Predict postop complications without sharing data | Multi-center, privacy-first |
more papers (25+ total) ...
- Risk-Specific Training Cohorts β class imbalance in surgical prediction β JAMA Surgery (2024)
- Transparent AI β explainable postop complication interface (2024)
- Global Contrastive Training β multimodal EHR + language supervision (2024)
- EHR Data Quality β scoping review β JMIR Med Inform (2024)
- Selective State Space Models β brain dysfunction across cohorts (2024)
- AKI Prediction β external validation for non-critical care (2024)
- ML + FHIR β scoping review on FHIR-native ML systems β JMIR Med Inform (2023)
- AI-Enhanced ICU β pervasive sensing in critical care (2023)
- Delirium from Ambient Signals β noise and light as ICU predictors (2023)
- Computable Phenotypes β brain dysfunction characterization (2023)
- AKI Multistate Analysis β kidney injury trajectories β Scientific Reports (2023)
- Acute Illness Phenotypes β temporal clustering (2023)
- Deep Interpolation Network β physiologic time series (2020)
I got tired of watching good models sit in Jupyter notebooks while clinicians still drown in paperwork. So now I build the infrastructure that gets AI out of the lab and into the clinic.
ββββββββββββββββββββββββββββββββββββββββββββ
β ClinicClaw (coming soon) β
β AI agents that do real clinical work β
β ambient docs Β· smart orders Β· prior authβ
ββββββββββββββββββββββββββββββββββββββββββββ€
β VERITAS β
β every agent action: checked, logged β
β deny-by-default Β· SHA-256 audit chain β
ββββββββββββββββββββββββββββββββββββββββββββ€
β HAVEN Protocol β
β patients own their data. period. β
β consent Β· provenance Β· value tracking β
ββββββββββββββββββββββββββββββββββββββββββββ€
β PSDL β
β clinical scenarios as code β
ββββββββββββββββββββββββββββββββββββββββββββ
HAVEN Protocol Β |
Before an AI agent touches a patient record, VERITAS checks the policy, logs the action, and verifies the output. Every. Single. Time.
The actual hospital system. Real LLM calls, real FHIR, real clinical workflows β all governed by VERITAS. AesculTwin Β |
impl Chester {
fn day_job(&self) -> &str {
"MS ECE @ UF Β· data scientist III Β· IC3 + PRISMAp"
}
fn what_i_actually_do(&self) -> Vec<&str> {
vec![
"wrangle EHR data into something models can eat",
"build ETL pipelines that don't break at 3am",
"embed models into systems clinicians actually use",
"make sure the whole thing is auditable and trustworthy",
]
}
fn tools(&self) -> Vec<&str> {
vec!["Rust", "Python", "TypeScript", "SQL", "FHIR R4"]
}
fn funded_by(&self) -> Vec<&str> {
vec!["NIH R01", "NIH R21", "NIH K01", "NIH OT2"]
}
}
