vaddify.dev
weekly briefs · live · Mon 8am PT

Build the AI stack
others write about.

~$ Production-grade AI agents on Azure.

Vaddify is the field journal of a Microsoft architect shipping AI agents on Azure. Patterns, runtimes, and the small ugly details no one tweets about.

Azure AI Foundry·Container Apps·MCP·Functions·Cosmos DB
// what's inside

For the people shipping AI on Azure.

Three audiences, one source of truth. Swipe through yours.

// for engineers

Ship agents, not slideware.

Production patterns for AI engineers — orchestration, tool-use, evals, and tracing — straight from systems running on Azure Foundry, Container Apps, and Cosmos.

Agent orchestration that holds up

Multi-agent crews, planner/executor splits, and authority boundaries that survive contact with real users.

Tools, MCP & function calling

Wiring agents to real systems via Model Context Protocol, function calling, and grounded retrieval.

Evaluation as code

Offline evals, golden sets, LLM-as-judge, and CI gates so quality isn't measured by vibes.

Observability & tracing

OpenTelemetry traces across prompts, tool calls, and model hops — debugging non-determinism with data.

Reference architectures on Azure

Foundry + Container Apps + Cosmos + ACS, wired correctly — fork-friendly repos with Bicep and CI.

What's actually new each week

Cutting through release-note noise to what changes your roadmap on Monday morning.

// for architects

Design choices that outlast the next model.

Reference architectures, patterns, and trade-offs for AI architects — RAG vs. fine-tune, single-agent vs. crew, PTU vs. PAYG — designed to outlast the next model release.

AI reference architectures

Opinionated, end-to-end blueprints for agentic apps, RAG, multi-agent crews, and enterprise copilots on Azure.

Patterns and anti-patterns

Field-tested calls: when to use RAG vs. fine-tune, single-agent vs. crew, one model vs. a model router.

Identity, networking, data boundaries

Managed identities, private endpoints, tenant isolation, regional residency — wired without compromise.

Designed for observability & eval

Tracing, evals, drift detection, and SLOs designed in from day one — not bolted on after launch.

Cost-aware architecture

Caching, distillation, model routing, PTU vs. PAYG — and the FinOps signals that prove the choice.

The architect's review checklist

A standing checklist for reviewing AI solutions against safety, scale, and exit-cost before they ship.

// for AI Centers of Excellence

Operating discipline for the org standardizing on AI.

An operating model for AI Centers of Excellence — use-case intake, shared platform, Responsible AI controls, FinOps, and enablement — for the org standardizing on Azure Foundry.

AI use-case prioritization

A repeatable intake and scoring framework to pick what to build, what to buy, and what to kill.

The shared AI platform

A central Foundry landing zone — models, identity, evals, guardrails — that teams want to use, not work around.

Responsible AI in practice

From policy to controls developers actually ship behind. Mapped to NIST AI RMF, EU AI Act, ISO 42001.

FinOps for AI

PTU vs. PAYG, chargeback, prompt caching — keeping the bill predictable as adoption scales.

Enablement that compounds

Talent ladders, internal certifications, and communities of practice that change developer behavior.

The CoE scorecard

A handful of leading and lagging metrics that tell leadership whether AI is actually working.

One email a week. Zero hype.

One brief every Monday — what shipped on Azure AI this week, with sources, plus the top frontier movements worth your attention. Free.

Subscribe