Model research
A complete workflow for converting spatial spectroscopy scans into reliable classifier inputs.
- Scan ingestion and spectral preprocessing
- Pixel sampling and feature extraction
- Multiple classifier variants and evaluation
From spectroscopy scans to a production AI platform.
BLK engineered the full path from ML research to secure on-prem operation—models, consistent inference, visualization, and deployment for the people using NomadX instruments in the field and lab.
- Client
- NomadX
- BLK scope
- ML research → training → production → visualization → deployment
- Environment
- Secure on-prem
The gap between the instrument and the operator.
NomadX's spectroscopy instruments produced rich spatial scan data, but operators had no practical way to run machine-learning analysis in the field or lab. The company needed more than a dashboard: it needed the models, a production inference system, and a secure local experience people could use without specialized ML tooling.
One production path from scan to decision.
BLK owned the connected system—not just one model or one interface. Each layer was engineered so research decisions remained consistent in the deployed operator experience.
Production inference
Production model bundles and an API layer that preserve the same preprocessing used during training.
- Exported model bundles
- Training-to-inference consistency
- FastAPI prediction services
Operator experience
A browser application that turns complex model output into an inspectable analysis workflow.
- Scan and pixel-level predictions
- Spatial heatmaps and spectral charts
- Confidence, class distributions, and history
Model operations
Controlled tooling for validating, monitoring, and updating models in on-site environments.
- Performance logging
- Validation imports
- Containerized on-prem deployment and updates
The hard part was preserving scientific intent in production.
The production API mirrors the preprocessing used during model training, keeping scan ingestion, spectral transformations, feature extraction, and classification aligned from notebook to deployment. Operators get a browser workflow; the underlying system keeps the modeling assumptions intact.
Operator experience
Upload scans, run inference, inspect confidence, spatial heatmaps, spectral charts, class distributions, history, and model selection.
Model operations
Performance logging, validation imports, controlled on-site updates, and a repeatable route from improved model to deployed system.
Complex ML became an operator workflow.
Raw scans became decisions
Rich instrument output moved through a repeatable ML pipeline into results operators could inspect and act on.
The browser replaced the notebook
Non-technical users could run classification and explore spatial and spectral evidence without working in research tooling.
Local by design
Data, inference, model selection, and updates remained on local infrastructure for secure field and lab operation.
Production AI, engineered as a complete system.
BLK connected scientific research, production software, operator experience, and deployment into one dependable system. The result was not a model demonstration—it was a working AI capability that NomadX could operate, validate, and improve in the environments where the analysis happens.