Machine Learning · Stealth Startup
PTSD Symptom Analysis Model
Problem
An AI-led PTSD symptom analysis product needed a model that could recognize meaningful language patterns from messy, real-world input.
Approach
Trained a language pattern-recognition model and built the full pipeline around it — data collection, cleaning, and training-set preparation from raw sources.
Impact
Became the core of the analysis pipeline; architecture and decisions were documented for reproducibility and onboarding.
Clinical signal in everyday language
PTSD symptom tracking is hard to scale clinically. The product's bet was that language patterns in user input carry detectable symptom signals — if a model could recognize them reliably, screening could reach people a clinic never would.
Owning the whole pipeline
This wasn't model work on a prepared dataset. The pipeline started at raw sources — web crawlers and manual collection — through cleaning and normalization, training/test set preparation, and finally model training for language pattern recognition. As founding technical lead, the data engineering and the model were the same job.
Evaluation where stakes are high
In a mental-health domain, a false positive isn't a rounding error. Evaluation prioritized precision and recall trade-offs over raw accuracy, with the false-positive rate treated as the gating metric.
Outcome
The model became the core of the AI-led symptom analysis pipeline. Architecture and pipeline decisions were documented for reproducibility, which made onboarding junior contributors possible — the system outlived any single person's memory of it.