TEDDY
A transformer trained from scratch to forecast first-occurrence diagnoses from longitudinal clinical event histories.
150M+raw clinical events tokenized
AI Engineer & ML Researcher
Johnathan Jia
I build and study machine-learning systems—from model development and rigorous evaluation to scientific computing and production engineering.
Computational scientist with deep biomedical experience, working across research questions, real-world data, and the systems around a model.
150M+clinical events processed
16Kreasoning rollouts analyzed
~99%scientific runtime reduction
Models + evaluation + systems
Selected work
Models and systems built around difficult data.
Work across transformer development, retrieval and ranking, interpretability, and scientific performance engineering.
Detectability of Doomed Reasoning Chains
Testing whether residual-stream representations can identify reasoning trajectories that are unlikely to recover before generation completes.
16Kverifier-graded rollouts
MentorMatcher
A production-oriented mentor search system where evaluation led from an unstable LLM judge to a fine-tuned cross-encoder reranker.
decontX Optimization
Profiling and redesigning a scientific Python bottleneck reduced a large single-cell workload from more than 12 hours to minutes.
~99%runtime reduction
Current research
Independent · Ongoing · 2026
Can a model know when its reasoning will fail?
I am testing whether residual-stream representations can predict a rollout’s eventual correctness before generation finishes—and where apparent trajectory signal is actually problem-difficulty signal.
Observed pooled AUC0.708with unresolved confounding and calibration questions
Read the research noteWorking across layers
Close to the problem, the data, the model, and the system.
My path through biology, quantitative science, computational genomics, and AI engineering shapes how I work: understand the scientific or operational constraint, build the model, measure its behavior, and make the surrounding system reliable.