StartupENGINEERING
An Applied Research Lab
A Discovery Lab for startups

Make engineering decisions with evidence, not vibes.

Every shipped product is layers of micro-decisions — architecture, frameworks, patterns, tools. The right answer in each one depends on your stage, team, and risk tolerance.

I work through one decision at a time: surface the source artifact, evaluate it in startup context, publish the analysis. The catalogs below are what's been evaluated so far.

Browse the catalogsSee the rubric
Run by Selva — Principal AI Architect (50M TPM Gen AI in prod), ex-founder of a 45-person SaaS, LinkedIn Top AI Voice 2024.
How the lab works

A decision artifact is any concrete thing a startup team might adopt: a reference architecture from AWS, a paper from Anthropic, a framework from Y Combinator, a vendor's tool.

  1. 1Pick a decision — “Should we adopt OAuth 2.1?”, “Is the AWS GenAI reference architecture right for a 10-person team?”
  2. 2Find the artifact — the paper, framework, vendor doc, or reference architecture that defines it.
  3. 3Evaluate in startup context, using a public rubric.
  4. 4Publish — verdict, trade-offs, what to adopt, what to skip.

Skills· 27

View all
Self-contained agent skill artifacts, scored against a public rubric for what actually works in a startup context.
Skill · Define

Idea Refine

Addy Osmani · updated 2026-04-20
Refines ideas iteratively. Refine ideas through structured divergent and convergent thinking. Use "idea-refine" or "ideate" to trigger.
Skill · Discover

Planning and Task Breakdown

Addy Osmani · updated 2026-04-20
Breaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible.
Skill · Discover

Spec Driven Development

Addy Osmani · updated 2026-04-20
Creates specs before coding. Use when starting a new project, feature, or significant change and no specification exists yet. Use when requirements are unclear, ambiguous, or only exist as a vague idea.

Patterns· 1

View all
My own thinking on recurring agent-engineering problems and the approaches that hold up under real pressure.
Pattern

MCP Gateway: a single control plane for agent-tool plumbing

Published 2026-04-27
Once a startup runs more than two MCP servers, security, logging, rate limiting, and policy start duplicating. The gateway pattern moves them into one hop in front of MCP.

Frameworks· 1

View all
Independent evaluations of conceptual frameworks published by researchers and AI labs — adopt, selective, or skip.
Framework · context-dependent

The 4-Layer Agent Evaluation Framework

Google · evaluated 2026-04-27
Google's four-layer evaluation framework for non-deterministic agents: components, trajectory, outcome, live system. Findings on what the layering contributes, where it under-specifies, and where each layer earns its eng-weeks.

Architectures· 3

View all
Reference architectures from cloud vendors and analysts, assessed for real startup fit — not enterprise fit.
Architecture · Azure

Baseline Microsoft Foundry Chat on Azure

Microsoft Learn · evaluated 2026-04-30
Microsoft's enterprise baseline (last reviewed April 2026) for a single, prompt-based agent chat on Microsoft Foundry: Azure App Service web app behind Application Gateway + WAF, agent runtime in Foundry Agent Service, RAG via Azure AI Search, conversation state in Cosmos DB, full VNet isolation with Private Link and Azure Firewall egress. Independent evaluation for startup fit.
Architecture · AWS

Building Agentic AI Foundation Solutions on AWS

AWS · evaluated 2026-04-29
AWS reference architecture (reviewed January 29, 2026) for an authenticated customer-support agent: Cognito at the edge, Amazon Bedrock AgentCore (Runtime, Identity, Memory, Gateway, Observability) wrapping a LangGraph agent, LiteLLM as a multi-provider model gateway, and a Bedrock Knowledge-Base RAG path over OpenSearch + S3, with Langfuse and CloudWatch for telemetry. Independent evaluation for startup fit.
Architecture · GCP

Multi-Agent AI System on Google Cloud

Google Cloud · evaluated 2026-04-29
Google Cloud's reference architecture (last reviewed September 16, 2025) for a coordinator-plus-subagents system: a Cloud Run chat frontend, an ADK-built coordinator routing to specialised subagents in sequential and iterative-refinement patterns, A2A for inter-agent calls, MCP for tool access, Vertex AI for inference, and Model Armor + VPC-SC for safety. Independent evaluation for startup fit.

Insights· 4

View all
Dated observations on where agent engineering is moving, with implications for the next 90 days.
Insight · ai-capability

The collapsing time horizon of code

Published 2026-04-27
Anthropic's capability curve compresses a year of engineering into hours. If your throughput is still measured in story points and sprints, you're benchmarking against the wrong baseline.
Insight · product-strategy

Evaluation is the new moat

Published 2026-04-20
Model access is commoditized. The defensible asset is your rubric, your golden dataset, and the discipline of running both against every change. That's where domain expertise compounds.
Insight · org-design

From authors to editors

Published 2026-04-15
When agents draft code, the engineer's value moves to shaping intent before, and judging output after. Promotion ladders, interview loops, and standups still measure 2019 work.

Tools· 0

View all
Evaluations of agent tools and developer products you might wire into a startup stack.
First tools coming soon.

Why this exists

Startups solve humanity's problems. Engineering is how they get built. I help the engineering get built right.

I spent 5 years founding a SaaS startup and scaling it to 45 people. Before that, 10+ years shipping enterprise systems at Brocade and HCL. Now I'm Principal Architect at an AI startup, running production LLM infrastructure at 50M TPM. I've been on every side of the table — the founder making impossible tradeoffs, the architect defending them, the engineer implementing them under pressure.

What startup teams lack isn't information. It's trusted judgment applied to their specific context. That's what this site is for.

  • 5 years as a startup founder
  • Principal Architect running 50M TPM Gen AI production
  • 18 years across four technology waves
  • LinkedIn Top AI Voice, Aug–Dec 2024

Stay in the loop

I publish evaluations as they're ready. Follow on LinkedIn for the fastest updates.

Follow on LinkedIn