What clients
have said.

Challenge

A packaging manufacturer had deployed a visual defect detection model using cloud inference, but the 400–600ms round-trip latency made the system unsuitable for inline quality checks on a moving production line. Cloud costs were also becoming substantial at production volumes.

Solution

Obsidium assessed the manufacturer's existing camera hardware, identified a viable edge device tier, and optimised the existing TensorFlow model for deployment on those devices without retraining. Monitoring pipelines were configured to report inference latency, detection confidence, and anomaly rates.

Results

Inference latency reduced to under 40ms on-device. Cloud inference costs eliminated. The model has operated in production for six months with two minor calibration adjustments triggered by the monitoring system. Overall defect detection accuracy maintained within the original model's specification.

Challenge

A healthcare research group needed 15,000 labelled clinical text samples for an NLP model targeting named entity recognition in Cantonese and English medical records. Previous annotation providers had produced inconsistent labels due to lack of medical domain guidance.

Solution

Obsidium developed annotation guidelines in a series of working sessions with the client's medical records team. Annotators worked from those co-developed guidelines under a two-pass QA process. A 500-sample pilot was completed first to validate the guideline interpretation before full-scale production.

Results

Delivered 15,000 labelled samples with 97.2% inter-annotator agreement on the core entity types. The pilot approach caught a definition ambiguity early, preventing a systematic labeling error across the full dataset. The guideline document was adopted internally by the client for ongoing annotation work.

Challenge

A multi-location retailer was using a demand forecasting model that had been deployed nine months earlier with no performance tracking. Seasonal pattern shifts had gone undetected, leading to notable over-ordering in two product categories during the Q3 period.

Solution

Obsidium implemented a monitoring layer that tracked prediction accuracy against actuals, flagged data distribution shifts relative to the original training period, and generated automated weekly summaries. Two dashboard views were built: a technical metrics view and a plain-language summary for the buying team.

Results

Within the first six weeks, the monitoring system flagged a distribution shift in one product category ahead of a demand spike, allowing the buying team to adjust their orders proactively. The client estimates this alone covered the monitoring setup cost. Two more drift events have been caught in subsequent months.