Send imaging referrals to the right administrative queue first

An imaging referral can sit idle because an administrative packet is incomplete, even when every system already holds part of the information. An AI workflow that writes a summary before identifying the missing item adds another step to the queue. Classifying the next administrative action can be more useful.

What this means for your business

For an outpatient imaging centre, the opportunity is fewer misrouted referrals and less staff time reopening the same packet. A focused router could identify which administrative workflow should examine a referral next. This is a proposal for paperwork handling, not a diagnostic system or clinical-priority model.

CLM-v0.1-8B scores supplied choices instead of writing an answer. Its two projection heads use a frozen Qwen3-8B encoder. That makes it a candidate for deciding which workflow should run next. It does not extract a complete record, execute a tool or prove that a document is correct. The model card and implementation are linked below.

A specific workflow: a lumbar MRI referral packet

Take a fictional referral for a lumbar MRI without contrast. The incoming fax bundle includes an order, insurance details and a cover sheet saying “authorization pending.” The practice's referral system already holds the patient identity match and ordering clinician record. The scheduling calendar is a separate system.

Local OCR converts the fax pages into text with page references. Deterministic checks identify whether a signed-order field and an authorization reference are present. CLM receives that compact administrative state and ranks approved next steps: order-document follow-up, authorization-status follow-up, scheduling-readiness review, duplicate-referral review, or manual review.

For this packet, authorization follow-up is the proposed route. That worker retrieves the existing authorization task, matches the requested service using the practice's maintained rules, and prepares an internal work item pointing to the relevant fax page. Staff verify the result. There is no reason to invoke a scheduling model, a patient-message drafter and an order-chasing model simply to discover that an authorization task already exists.

Later, a new document may change the state and trigger a fresh classification. The router does not decide medical necessity, authorize a procedure or infer urgency from symptoms. Clinical urgency follows the centre's established clinical process regardless of the administrative route. Poor OCR or conflicting identifiers go to staff before any automated handoff.

A compact on-site deployment

The lightweight part is the scoring layer. The full service still needs the Qwen3-8B encoder. As a sizing calculation, eight billion weights at two bytes each occupy roughly 16 GB before runtime memory, activations and serving overhead. A single Linux workstation with a supported 24–32 GB GPU is a reasonable starting configuration to evaluate for short inputs and modest concurrency, not a tested minimum or a throughput promise.

The reference implementation pairs a vLLM pooling service with the CLM API. Keep decision inputs short, precompute reusable candidate embeddings, and batch arrivals within an acceptable waiting window. The repository's example caps inputs at 2,048 tokens; split longer documents into meaningful sections and retain page references instead of silently truncating evidence. One encoder can serve several department-specific sets of choices.

Four-bit weights would be roughly 4 GB in raw weight storage by arithmetic, but that is not the total memory requirement. Quantization changes embeddings and may change rankings. Validate the exact encoder, pooling method, precision and heads together before buying smaller hardware. The supplied heads cannot simply be attached to any small embedding model. A tiny CPU appliance or a Mac mini is not a verified drop-in deployment for this reference stack.

Keep the administrative AI next to the referral system

When a provider can maintain the infrastructure, on-site hosting is particularly useful because referral text, identifiers and routing logs need not be sent to an external inference service. Put OCR, classification and any drafting model on the controlled network. Limit service accounts to the referral queues they support, and retain document-level access controls when staff open the output.

A local deployment does not establish regulatory compliance by itself. The practical design includes audit access, retention settings, encrypted storage, patching and a fallback queue when the workstation is unavailable. Cloud-connected source systems still have their own data flows; document those separately instead of calling the entire workflow offline.

Measure completed administrative work

Evaluate on de-identified historical packets with staff-approved labels, including multiple referrals in one fax and conflicting authorization notes. Track wrong-queue assignments, manual-review rate, repeated model calls and staff time to a verified next action. Measure whether the handoff improves without adding missed follow-ups. Do not report clinical outcomes or patient wait-time improvements from a classifier-only benchmark.

The release is an early candidate-ranking model. Published agent benchmarks do not establish accuracy or speed on this workflow. Scores are relative to the alternatives supplied, not calibrated probabilities that a business decision is correct. Include an explicit manual-review choice, then set abstention rules using held-out examples; adding that choice alone does not make the system safe.