Open-Vocabulary Spectral Intelligence for Through-Package Pharmaceutical Verification
Closed-set classifiers recognize known fakes. PELLUCID reads unknown chemistry — describing and quantifying never-catalogued adulterants through sealed packaging, in the language of chemistry itself.
Today's detection asks one question: "is this one of the N substances I was trained on?" A counterfeiter only needs to be substance N+1. Novel adulterants are discovered in the market — after patients are harmed — not at the port of entry.
Physics-informed neural inversion of light transport through real packaging — foil, blister, amber glass. Packaging is a causal intervention, not noise. Quantitative chemistry out.
Recovered chemistry maps into a shared embedding with the written corpus of chemical knowledge. Describes never-catalogued adulterants — open-vocabulary description, not open-set rejection.
A generative counterfeit twin manufactures threats that don't exist yet and red-teams the system before fielding. A validation paradigm with no published prior art.
Calibrated risk, quantified uncertainty, plain-language narrative, spectral evidence trail — built for FDA auditability at port throughput.
PELLUCID proposes into TA2 DETECT as a universal decoding layer: one intelligence stack that ingests Raman, NIR, hyperspectral, terahertz, photoacoustic, or fused streams — and makes every TA1 sensor's fielded accuracy, throughput, and regulatory story stronger. We are actively forming teams for the FASTPASS program.
"We bring the analytics. Your sensor is the star."
The interactive walkthrough runs the full pipeline on sealed-shipment scenarios: acquisition through packaging, physics inversion, an open-vocabulary AI read, and a calibrated verdict certificate. The adversary loop invents a new counterfeit on demand and the system reads it anyway.