A financial crimes agent on AMD Instinct™ MI350P.Triage, evidence, narrative and QA for every alert on one LIQID UltraStack server, with SAR data inside your walls.
Route alerts into a growing queue, or investigate every one on infrastructure you control.
Alert volumes grow faster than investigator teams can hire.
Investigators spend most of their time gathering evidence across systems.
What U.S. financial institutions spend on AML operations.
Agentic workflows use 10–100× more tokens per task than one AI call.
Customer and SAR data can’t go to a hosted model.
Pinned, self-hosted models are easier to take through model risk review.
One 10-GPU LIQID UltraStack server runs every step after an alert lands. A small model routes it, a 70B model drafts the cited narrative, and a reasoning model checks it against known typologies.
Llama 3.1 8B replicas score and route every alert.
Llama 3.3 70B, one model per GPU, gathers evidence and drafts the SAR narrative.
gpt-oss-120b tests each draft against known typologies.
No cluster: three models, ten GPUs, one server, and the next job lands on the same host.
Add GPUs and a new job, not a new cluster.
Triage, evidence, narrative and QA for every alert.
4 × Llama 3.1 8B triage · 4 × Llama 3.3 70B evidence and narrative · 2 × gpt-oss-120b typology and QA
Head of Financial Crimes / BSA Officer. Alert backlog, exam findings, investigator attrition.
A department’s jobs on one server: triage, loan and KYC extraction, complaints, credit memos and QA.
4 × Llama 3.1 8B triage · 6 × Llama 3.3 70B extraction · 2 × Mistral Small 3.2 complaints · 4 × Llama 3.3 70B credit memos · 4 × gpt-oss-120b review and QA
COO / Head of Operations. Several pilots competing for GPUs; API bills growing.
Every line of business on one host, one model per GPU.
10 × financial crimes · 8 × lending and KYC · 6 × customer service and complaints · 6 × gpt-oss-120b research and developer assistant
CIO / Head of AI Platform. Consolidating departmental AI onto one platform.
Two questions settle the size of the opportunity: how many alerts sit untouched today, and how long does each one take to close?
*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.
Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models. Outputs support, and do not replace, investigator decisions and SAR filing. AML spend: FIS, May 2026. Token data: Futurum Research, Sep 2026.