choice,picked,why
Reasoning model,"GLM 5.3, filed at 753B parameters, FP8 (753 GB of weights), GLM-5.3 bespoke license, served on vLLM on one node in the operator's own facility in the United States","The model the requirement names; its license permits commercial use, fine tuning and redistribution, so the operator holds its own copy with no per token obligation; carries the triage and recovery drafting, the command center's cited conversational interface and ontology maintenance"
Site tier parsing model,"Qwen3.8 27B, dense, Apache-2.0, served at FP8 (about 31 GB of weights) on one 48 GB PCIe card in an air cooled server","High volume thread and note parsing over years of archived email and chat and the live threads, a throughput workload kept off the frontier node so coordinator sessions keep their concurrency; about 13 GB of KV ceiling on the card"
Embedding model,"Qwen3-Embedding 0.6B, Apache-2.0, about 1.2 GB at BF16 and about 0.6 GB at 8 bit, sharing the site tier server, 32K context window","Embeddings over threads, notes and resolutions so retrieval finds the past exception that matches this week's; permission aware, returning only what the asking coordinator could already open"
Frontier hardware,"141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB)","Holds 904 GB (753 GB of FP8 weights times the 1.2 planning factor) with about 280 GB of KV cache ceiling at the 0.92 utilization the sizing rule allows; about 15 kW of rack power per node with margin, 10 RU, 10 to 35 C inlet"
Frontier footprint sizing rule,"Frontier Model GPU Footprint Sizing Rule (vLLM, SGLang): parameters times one byte at FP8 times 1.2",Turns filed parameter counts into a memory requirement before any workload is placed
Node count rule,GLM 5.3 node count by GPU class,"One node of eight 141 GB GPUs; two nodes of the 80 GB class or BF16 serving (about 1.5 TB) rejected because they add a second node joined over an interconnect"
Site tier hardware,"48 GB PCIe card class, one card, in an air cooled PCIe server the facility's server room already takes",Fits the parsing model's FP8 weights with KV ceiling to spare and carries the embedding model beside it without a second server
Serving runtime,vLLM,Serves the FP8 weights within the KV cache ceiling the sizing reserves
Registry,MLflow Model Registry on site,Pinned model images; serving restarts from pinned images with a rollback exercised once under observation; the desk never runs on an unobserved weight change
Event backbone,Apache Kafka (KRaft mode),"Ordering by offsets and per stream sequence numbers with each source's clock offset recorded, at least once delivery with idempotent projection, buffering in the retained log, replication inside the facility"
Time series store,InfluxDB 3 Core,Tracking telemetry under the object model
Orchestration,Kubernetes,Containers for the services and the serving tier inside the facility
Identity,Keycloak,Federated identity for coordinators and permission aware retrieval scoped by each shipper's data terms
Observability,Prometheus,Observes the running system
Sensing,"None; the existing carrier tracking and visibility feeds are read","Trucks are seen through feeds the operator already owns; no camera, positioning hardware or edge box is built"
Pattern: memory,"Memory loops over confirmed outcomes, nothing trained, fine tuned or distilled into weights","The requirement's own words: better every week without training a model on the fly; memory is auditable where weights are not"
Pattern: approval,Every draft confirmed by a named coordinator; one gated write path into the scheduling tool,"Confirmations logged as ground truth, with known answer probes and weekly sampling against rubber stamping"
Pattern: baseline,The pilot lane group's exception measures are recorded before any coordinator sees a draft,"Improvement is measured against evidence, not impression; the phase 1 gate reads time to resolve and repeat exceptions against that baseline"
Pattern: boundary,"The operator's own hardware in its own facility in the United States; no inference call, retrieval or draft leaves the building; only approved carrier messages and shipper notices cross","Shipper and carrier data never leave the operator's hardware, open weights only, no third party model service"