Customer Service Automation · Document Processing · Sales Automation
MISA needed to move from expensive, externally hosted LLM APIs to a self-hosted Vietnamese AI stack that could protect sensitive data, reduce latency and operating cost, and support many concurrent public-service and enterprise automation workloads.
Reported impact
MISA cut fine-tuning time by 4x, increased throughput from 1,200 to 4,500 TPS per model, saved about $200,000 per month, automated recruiting workflows in minutes with up to 90 percent HR time savings, and now uses the platform to support more than 450,000 organizations and over 1.1 million daily LLM requests.
MISA built an on-premises AI factory on NVIDIA infrastructure so it could own the full lifecycle of Vietnamese domain models and AI agents. The core cluster uses eight NVIDIA DGX H200 systems linked with NVIDIA Quantum InfiniBand networking, while NVIDIA A40 GPUs handle smaller specialized tasks such as eKYC and router LLMs. NVIDIA Run:ai virtualizes the cluster and enables fractional GPU sharing, which increased concurrent experimentation dramatically. On top of that hardware, MISA uses NVIDIA NeMo for domain-specific tuning, NVIDIA Dynamo and NVIDIA NIM for production-scale serving, and FP8 quantization to raise inference throughput for real-time agent interactions. The AI factory powers more than 50 prebuilt agents and hundreds of templates, including MISA AVA, MISA AMIS OneAI, Support, Mail, and Recruitment assistants, so accounting, tax, HR, sales, and public-administration workflows can run securely on self-hosted infrastructure instead of external APIs.
Business and public-service data enter MISA’s self-hosted AI factory, DGX H200 clusters and NeMo fine-tune domain models, Dynamo and NIM serve those models to AVA and other agents, and the outputs automate accounting, HR, sales, support, and public-administration tasks inside customer workflows.
By building an on-premise AI factory based on NVIDIA technologies, MISA aims to gain full technological ownership, reduce high operational costs, and deploy domain-specific AI agents capable of automating complex public services and enterprise workflows with low latency and high data security.
By deploying an end-to-end NVIDIA-accelerated full-stack architecture ... MISA reduced fine-tuning times by 4x and achieved $200,000 in monthly cost savings.
By leveraging FP8 quantization, inference throughput has increased from 1,200 to 4,500 TPS per model.
MISA deployed an on-premise cluster featuring 8 NVIDIA DGX™ H200 systems connected with NVIDIA Quantum InfiniBand networking.
With AI, the entire workflow—from CV intake and screening to evaluation and data extraction—is 100% automated within minutes, saving up to 90% of HR processing time.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Sep 1, 2026 Onetera Onetera needed to help more than 90 municipalities turn fragmented regulations, zoning maps, permit rules, and legacy documents into fast resident-facing services without imposing heavy IT overhead or expensive manual… 1.7x increase in data throughput GPT OSS 20B · Government Intelligence Agent · LangGraph NVIDIA Sample Jan 17, 2025 FPT Smart Cloud FPT Smart Cloud needed more natural, multilingual virtual assistants and more scalable employee upskilling systems, but high speech-synthesis training costs, long processing times, and the need for personalized… 100X improvement in speech synthesis model output in dimension and aptitude across 18 languages FPT AI Engage · FPT AI Mentor · Kubernetes 
In-house Sample Apr 8, 2025 Bank of America Bank of America needed to improve employee productivity and client service at scale across a global workforce of 213,000 people, while reducing routine support demand, accelerating employee access to information,… 20 million clients now actively using the virtual assistant Erica · Erica for Employees · PRIVATE BANK 