AI Atlas - Daily updates on real-world AI deployments worldwide

Tencent Cloud describes building CloudQ, an AI-native SRE agent that combines an operational knowledge graph with multi-agent orchestration for architecture visualization, impact analysis, root-cause investigation, change-risk control, inspection, capacity planning, and FinOps. The graph joins infrastructure topology, observability signals, change events, and operational knowledge. A Harness layer adds bounded agent execution, sandbox isolation, multi-cloud adapters, and online/offline evaluation with regression gates. The InfoQ article is an edited conference-talk report; its incident walkthroughs are illustrative and it provides no independently measured customer outcomes.

Revolut, the London-based digital bank, self-hosts an open-weight GLM coding model on 56 NVIDIA B200 GPUs to give about 2,000 engineers flat-cost, unmetered AI coding. Its AI operations lead reports a 60% cut in coding time and 70% in lead time after the switch in April 2026, a blended cost of about $0.30 per million tokens, roughly one tenth of its previous frontier-lab spend, and an 88% latency improvement from inference tuning on the same hardware. A sandbox with operation allow-lists keeps a human in the loop for EU banking supervision.

InfoQ reports Xiaohongshu scaled its OpenClaw-based Seal from a three-day demo to all-staff use in about a month. It describes cloud Kubernetes and Nex microVM sandboxes. The team reported Self-GC lowered daily token use 10–15% (about 20% at peak); Auto routing lowered model costs about 69%.

The AI-IR lab at the University of Texas Medical Branch reimplemented 21 pretrained medical imaging models, including nnU-Net, TotalSegmentator, MedSAM2 and SAM 3, natively in MLX for Apple Silicon. Benchmarked against the original implementations on an Intel CPU and an NVIDIA RTX 4090 across public cases from 11 imaging datasets, the models matched outputs in 211 of 212 comparisons, and 19 of the 21 ran in under 16 GB of memory on a 64 GB Apple M1 Max, with a memory-saving setting cutting memory use by a median 49% at a median 13% processing-time cost.

PetroChina's Research Institute of Petroleum Exploration & Development built a shale oil large language model on Qwen2.5-72B and Qwen3-32B using a 21-billion-token domain corpus, QLoRA fine-tuning with dynamic context parallelism and a reflective RAG loop, reaching 86.35 on objective and 75.54 on subjective shale oil tests versus 77.32 and 12.12 for GPT-4o. A follow-on agent built on the Model Context Protocol orchestrates 23 geoscience tools for seismic processing and interpretation and scored 94.377 on a four-part tool-calling benchmark, above ChatGPT-5.4 and Kimi k2.6.

Slack rebuilt how it serves the large language models behind Slack AI in four phases: models hosted in an escrow VPC on AWS SageMaker, a zero-incident migration to Amazon Bedrock, a hybrid of provisioned and on-demand capacity with spillover, and from early 2026 a multi-cloud setup adding Google Cloud Vertex AI. A routing layer picks a model per feature, with backup models and circuit breakers, lifting complex-reasoning quality about 10% and cutting latency about 67% on low-token workloads.

Johns Hopkins researchers released Merlin-Cancer-Super, a model that detects and segments tumors in nine organs on CT scans. It is trained with R-Super, which uses an LLM to extract tumor count, size and location from radiology reports and turns them into per-voxel training signals, on Merlin Plus, a dataset built with Stanford from 25,494 CT scans. The authors report it outperforms leading public models and vision-language models.

Morpheus Studios produced a full three-minute action sequence for the SBS drama "Agent Kim: Reactivated" entirely in its node-based AI video platform AICRON, with no live-action footage: about 500 concept images, 1,000 AI images and 2,000+ video clips yielded 289 broadcast cuts, including AI close-ups of the lead actor made with his consent. The series became 2026's highest-rated Korean miniseries.

Boston Medical Data runs the open-source DocRouter platform on-premises to turn mixed fax and PDF peri-operative packets of up to about 500 pages into patient-level data for anesthesia revenue cycle management. A workflow splits each packet, runs Textract OCR and an LLM page classifier in parallel per page, groups pages by MRN or name and date of birth, and extracts billing fields, with unassigned pages sent to human review. DocRouter reports 175,000+ documents processed at 99.9% extraction precision.

The French government's DINUM ran a 10,000-licence pilot of its sovereign Assistant IA, and the DITP evaluation found document synthesis fell from 33 to 28 minutes with quality scores rising from 14.15 to 15.43 out of 20.

SaM Solutions runs a self-hosted Gemma 3 model inside its Umbraco CMS to translate 1,300-plus internal pages, cutting an estimated 2.5 weeks of manual work to a few hours with a conditional publish-and-review workflow.

Indian digital lender Kissht runs an Arrowhead voice AI agent, in Hindi and English, on two inbound support queues: loan-rejection queries and mandate/NOC cancellations. It is integrated with Kissht's dialer and CRM, hands any mention of the RBI to a human within the same turn, and runs on Arrowhead's own small language model at about 500 ms turn latency. Within two months it handled 40% of all inbound calls, resolving about 88% without a human at roughly 30% lower handle time.

Flock, which insures commercial vehicle fleets, built a pipeline that scores telematics crash files with an XGBoost model on 42 engineered features, then has two Claude Haiku agents on Amazon Bedrock check data quality and the vehicle's claim history before a verified crash is broadcast as an automatic first notification of loss. Every prediction is archived for audit and retraining.

Travelers built AI Claim Assistant, a fully autonomous voice agent on OpenAI's Realtime API and frontier models that handles first notice of loss for auto property damage: it answers policy questions, gathers details and submits the claim. After launching in eight states it went countrywide within two months, and 85-90% of customers who use it finish filing through AI.

Sony Bank and Fujitsu applied Anthropic's Claude on Amazon Bedrock, with Claude Code and MCP-connected agents, to the development of Sony Bank's production core banking system from September 2025. By July 2026 the development period from basic design to integration testing was 30% shorter and man-hours 40% lower, with humans keeping final judgement and quality assurance.

Starling Bank moved its data warehouse to BigQuery and its machine learning to Vertex AI, then built Gemini-based tools: Scam Intelligence checks marketplace ads for scam signs and has raised payment cancellations by 300%, call-transcript summaries for staff save 8,000 hours a month, and a Gemini help centre cut customer service referrals by 50%.

AZORG says it is the first hospital in the world to use a fully integrated, AI-driven approach for chest pain patients, built with PMcardio. In 2025 the platform analysed 32,807 ECGs from 10,854 patients and detected more than 500 heart attacks, with an average 48 minutes saved from diagnosis to treatment.

Swiss digital insurer Simpego and AWS partner copebit replaced manual claims data entry with a serverless pipeline that sends each uploaded document to Anthropic's Claude Opus on Amazon Bedrock and validates the output with schema and VAT arithmetic checks. Built in two months, it runs at about 30 seconds per document, with 95% needing no manual intervention.

State Farm samples real conversations from its production AI agents for chat center representatives and scores them with Claude Sonnet 4.5 on Amazon Bedrock through DeepEval. A stricter custom faithfulness metric exposed unsupported inferences in a billing agent, and a prompt fix raised its score 8%. In validation against human reviewers, 78% of judge scores were within 3 points.

Zurich Commercial Insurance uses Cytora as a headless intelligence layer that turns unstructured, multilingual submissions into decision-ready risks for its Z-Horizon+ underwriting workbench. The page describes a rollout from seven countries in late 2025 to four more in Q1 2026 and a self-serve model from Q2 2026, with 95%+ accuracy and a goal of 15-minute triage.