Allied Vision has introduced Alecs, an edge-AI platform for VSWIR (visible to short-wave infrared) machine-vision applications, MVPro Media reported.

The announcement extends on-camera inference into the short-wave infrared band, where silicon CMOS sensors lose sensitivity and InGaAs (Indium Gallium Arsenide) detectors take over.

VSWIR spans roughly 400 nm to 1700 nm. The window includes silicon-wafer transmission cutoffs near 1100 nm, moisture absorption bands around 1450 nm and 1940 nm, and material-specific features used in food-quality grading, semiconductor metrology, pharmaceutical verification, and recycling-stream sorting.

Why VSWIR needs its own inference layer

Short-wave infrared inspection pipelines face a data-rate problem visible-light systems do not.

Raw InGaAs frames at 12 or 16 bits per pixel, delivered over CoaXPress or 10 GigE, often exceed the host PC's preprocessing budget before inference begins. Pushing those frames to a server-side CNN adds round-trip latency and a second failure domain.

On-camera inference collapses that loop. A CNN running on an FPGA or SoC co-processor returns classifications within the camera's exposure interval, so a sorter acts on the same frame the sensor just captured.

MVPro Media's report does not specify three values that decide procurement:

• Accelerator silicon (FPGA, SoC, or dedicated NPU)
• End-to-end inference latency at the application's working frame rate
• Supported model interchange (ONNX, TensorRT, OpenVINO)

Without those numbers, an instrumentation buyer cannot tell whether Alecs fits an inline sorter at 100 fps or only a slower offline review station.

The calibration problem Alecs must solve

VSWIR cameras ship with non-uniformity correction (NUC), bad-pixel maps, and radiometric calibration coefficients stored in on-board memory.

These compensations handle pixel-to-pixel gain variation, dark-current offsets, and lens vignetting. An embedded inference layer that intercepts the pixel stream before these corrections run will distort classification.

Alecs must execute after NUC and bad-pixel substitution, or the model trains on artefacts production does not reproduce. The report does not state where in the image-processing chain Alecs sits.

What thermal envelope does the housing provide?

InGaAs dark current roughly doubles every 7-10 °C, so the sensor's noise floor tracks housing temperature.

Adding a neural-network accelerator behind the sensor package raises heat dissipation. A poorly cooled TO39 or TO8 housing can push the sensor above its specified operating range. The report does not enumerate the Alecs power budget.

What compliance question does the integration raise?

Inline VSWIR inspection in pharmaceutical and food lines operates under GAMP 5 validation frameworks and FDA 21 CFR Part 11 for electronic records.

Embedding a neural network behind the sensor introduces a new audit boundary. Model versioning, retraining cadence, and inference-log retention become part of the validation dossier.

Reviewers will ask how Allied Vision documents those elements and whether Alecs supports deterministic rollback to a prior approved model version.

MVPro Media's report does not state release date, pricing, supported Alvium models, or qualifying OEM partners. Until those appear in a datasheet revision or vendor FAQ, an instrumentation team should treat Alecs as an architecture disclosure rather than a procurement-ready product.