Research & playbooks
for shipping AI in Asia.
Frameworks we use in client engagements, plus original research on AI adoption across the markets we operate in. No hype, no rehashed Western reports.
APAC LLM Observability Guide 2026: Arize Phoenix, AgentOps, and Lunary
A practitioner guide for APAC AI engineering teams implementing LLM observability in 2026 — covering Arize Phoenix as an open-source OTel-based LLM tracing platform providing span-level visibility into RAG retrieval quality and agent workflows with automated hallucination detection and relevance scoring for CI/CD quality gates; AgentOps as an agent-focused observability platform with one-line framework instrumentation for LangChain, AutoGen, and CrewAI providing session-level replay, step-by-step agent action tracing, and real-time cost anomaly detection; and Lunary as a lightweight open-source LLM logging platform capturing prompts, responses, and costs across production deployments with user feedback collection for CSAT tracking and self-hosted PostgreSQL deployment for APAC data sovereignty requirements.
APAC LLM Inference API Guide 2026: OpenRouter, Fireworks AI, and Together AI
A practitioner guide for APAC AI engineering teams choosing managed LLM inference APIs in 2026 — covering OpenRouter as a unified API marketplace routing requests across 100+ models including GPT-4o, Claude, Llama, and Qwen with real-time per-token cost comparison and automatic fallback routing; Fireworks AI as a high-performance inference platform delivering sub-100ms time-to-first-token for open-source models via custom CUDA optimization with LoRA fine-tuning and dedicated hosted endpoints; and Together AI as an open-source LLM cloud providing access to 50+ models including Qwen 2.5 and DeepSeek via competitive per-token pricing with LoRA fine-tuning and dedicated GPU instances for APAC domain-specific model customization.
APAC Log Management Guide 2026: Grafana Alloy, OpenObserve, and Parseable
A practitioner guide for APAC DevOps and platform engineering teams modernizing log management infrastructure in 2026 — covering Grafana Alloy as an OpenTelemetry-native telemetry collector replacing Grafana Agent with a declarative component model that collects logs, metrics, and traces from APAC Kubernetes DaemonSets and ships to any OTel-compatible backend; OpenObserve as a Rust-native Elasticsearch-compatible platform using S3/GCS/MinIO object storage to achieve 140x lower log storage costs than Elasticsearch for APAC high-volume log workloads with unified logs, metrics, and traces in one platform; and Parseable as a lightweight Rust-native Parquet-based log ingestion and search engine for APAC edge deployments and resource-constrained Kubernetes clusters requiring minimal infrastructure footprint with 80% storage reduction versus Elasticsearch JSON indices.
APAC MCP and AI Gateway Guide 2026: FastMCP, MCP Inspector, and Portkey
A practitioner guide for APAC AI engineering teams building MCP server infrastructure and LLM gateway routing in 2026 — covering FastMCP as a Python decorator-based framework for building Model Context Protocol servers that expose APAC internal data sources and tools to Claude and MCP-compatible AI systems using @mcp.tool(), @mcp.resource(), and @mcp.prompt() decorators with stdio and SSE transport options; MCP Inspector as the official Anthropic interactive debugging tool for validating MCP server schemas, manually executing tool calls, and reproducing AI client behavior without adding Claude to the debug loop; and Portkey as an AI gateway platform providing multi-provider LLM routing with automatic fallbacks across OpenAI, Anthropic, and Azure OpenAI, semantic caching for repetitive APAC queries, prompt versioning and A/B testing, and per-model cost observability for APAC production LLM applications.
APAC AI Browser Automation Guide 2026: Stagehand, browser-use, and Browserbase
A practitioner guide for APAC AI engineering teams building browser automation and web agent workflows in 2026 — covering Stagehand as an open-source AI browser automation framework combining Playwright with LLM natural language act, extract, and observe primitives that identify DOM elements visually rather than by CSS selectors for resilient APAC web automation; browser-use as a Python library enabling LLM agents to control browsers via screenshot-based reasoning with multi-tab session management for APAC research agents integrated into LangChain and PydanticAI workflows; and Browserbase as managed cloud browser infrastructure providing scalable headless Chromium sessions with full session replay for debugging, residential proxy rotation for anti-bot evasion, and APAC geolocation options including Singapore, Hong Kong, and Tokyo for region-locked content access.
APAC LLM Framework Guide 2026: Semantic Kernel, DSPy, and Guidance Compared
A practitioner guide for APAC AI engineering teams selecting specialized LLM frameworks in 2026 — covering Microsoft Semantic Kernel for enterprise .NET and Python AI orchestration with typed plugin functions, SK Planner for automatic plugin selection, and Azure OpenAI integration for APAC data residency; Stanford DSPy for programmatic LLM pipeline optimization using declarative module signatures and MIPRO optimizers that automatically tune prompts and few-shot examples from APAC labeled datasets rather than manual prompt engineering; and Microsoft Guidance for token-level constrained generation that physically prevents LLMs from producing tokens violating APAC JSON schemas, enum values, or regex patterns — eliminating structured extraction parse errors in APAC document processing pipelines.
APAC Vector Database Extended Guide 2026: Vespa, LanceDB, and turbopuffer
A practitioner guide for APAC AI engineering teams selecting specialized vector database architectures in 2026 — covering Vespa as an enterprise search and recommendation engine with hybrid BM25 plus dense vector retrieval in a single query, real-time millisecond-fresh indexing at billions of APAC documents, and multi-phase ranking expressions for complex business rules; LanceDB as an embedded serverless vector store using the Lance columnar file format that colocates APAC vector embeddings with raw data in PyArrow-native tables on local disk or S3 without a separate database process; and turbopuffer as a pay-per-query serverless vector database backed by cloud object storage that loads relevant APAC index partitions on-demand for sub-second search latency without idle infrastructure cost for large collections with bursty query patterns.
APAC AI Agent Frameworks Guide 2026: AutoGen, PydanticAI, and smolagents Compared
A practitioner guide for APAC AI engineering teams evaluating next-generation agent frameworks in 2026 — covering Microsoft AutoGen for multi-agent conversation systems where specialized LLM-powered agents with defined roles collaborate through GroupChat orchestration with sandboxed Python code execution and human-in-the-loop UserProxy agents; PydanticAI for production Python LLM applications using Pydantic-validated structured outputs, typed dependency injection for testable APAC agent code, and TestModel for CI-friendly unit testing without LLM API calls; and HuggingFace smolagents for lightweight code-writing CodeAgents that generate Python to call tools using local open-source models without OpenAI API dependency for APAC data sovereignty requirements.
APAC eBPF Kubernetes Observability Guide 2026: Hubble, Pixie, and groundcover
A practitioner guide for APAC platform engineering teams adopting eBPF-powered Kubernetes observability in 2026 — covering Hubble as the Cilium ecosystem network observability layer using eBPF kernel probes to provide real-time service dependency maps, L7 flow inspection, and DNS query visibility for APAC clusters running Cilium CNI without sidecar proxies; Pixie as a CNCF sandbox auto-instrumentation platform that deploys as a Kubernetes DaemonSet and captures HTTP request traces, PostgreSQL query text, and DNS flows via eBPF in minutes without modifying APAC application code; and groundcover as an eBPF-native APM platform correlating auto-collected application traces with Kubernetes pod resource metrics in a unified UI compatible with OpenTelemetry SDK enrichment for APAC business context.
APAC AI Coding Assistants Guide 2026: Cursor, Windsurf, and Sourcegraph Cody Compared
A practitioner guide for APAC engineering teams evaluating next-generation AI coding assistants in 2026 — covering Cursor as an AI-first VS Code fork with full repository vector indexing enabling codebase-aware completions and multi-file Composer edits that plan changes across routes, controllers, and tests from a single prompt; Windsurf as an AI-native IDE by Codeium with the Cascade agentic AI that autonomously plans and executes multi-step coding tasks in Write mode, reading the APAC codebase and applying changes across multiple files with progress visibility; and Sourcegraph Cody for enterprise teams needing AI assistance grounded in multi-repository Sourcegraph code intelligence with fully self-hosted deployment using Azure OpenAI or AWS Bedrock for APAC data sovereignty in regulated industries.
APAC ML Model Monitoring Guide 2026: Evidently, whylogs, and Fiddler for Production ML Health
A practitioner guide for APAC ML engineering teams implementing production model monitoring in 2026 — covering Evidently AI for open-source drift detection using Kolmogorov-Smirnov and chi-square statistical tests generating HTML reports and pipeline-integrated test suites that block APAC batch inference when data quality thresholds are exceeded; whylogs for privacy-preserving statistical data profiling using sketch-based approximate histograms and cardinality summaries that enable drift comparison without storing raw APAC data externally; and Fiddler AI for enterprise ML observability combining SHAP/LIME per-prediction explainability with performance monitoring as delayed APAC ground truth labels arrive and fairness metric tracking across demographic segments for regulated industry compliance.
APAC ML Model Serving Guide 2026: Triton, Ray Serve, and MLflow for Production Inference
A practitioner guide for APAC ML platform and MLOps teams implementing production model serving in 2026 — covering NVIDIA Triton Inference Server for GPU-optimized inference with dynamic batching and TensorRT optimization across PyTorch, TensorFlow, ONNX, and Python backends on APAC GPU Kubernetes clusters; Ray Serve for Python-native LLM pipeline serving with composable deployments routing APAC requests to vLLM backend workers and per-deployment autoscaling on Ray distributed clusters; and MLflow Models for registry-integrated model packaging with one-command REST API serving and Docker container export for APAC Kubernetes deployment linked to training run provenance.
Want these in your inbox?
Subscribe to the RSS feed or talk to us about a research engagement on a topic specific to your firm.