2025
Multimodal Spectral Drug Detection
Deep learning models over spectral scans for a portable field detection device, hardened against sensor noise and distribution shift.
Built at Lightsense Technology for a handheld drug detection instrument. Laboratory accuracy is the easy half of this problem; the hard half is that the device goes into the field, where the sensor is noisier, the operator is less careful, and the compounds are cut with things the training set never saw.
What I built
- Multimodal models over spectral scans. Benchmarked architectures fusing multiple excitation channels against single-channel baselines on target compound detection.
- Adversarial evaluation protocols. Rather than a random test split, I built evaluation suites that deliberately induce distribution shift and simulated sensor noise, so the reported number reflects field conditions rather than bench conditions.
- UV-LED excitation optimization. Used the predictive models in the loop to select excitation configurations, improving detection sensitivity on target compounds — a case where the model informs the hardware, not just the output.
- A LoRA-adapted domain LLM for spectrometry question answering, evaluated specifically on whether its answers stayed grounded in calibration and drift documentation rather than confidently improvising.
Why the evaluation design mattered most
The adversarial protocol changed which model we shipped. The architecture with the best clean-split accuracy was not the one that held up under simulated sensor degradation, and picking on the clean number would have shipped a model that looked good in the report and failed on arrival.