L1 Artificial Intelligence Models, retrieval, agents, governance
OpenAI Anthropic Google Gemini Mistral DeepSeek OSS Llama / Qwen pgvector Qdrant Pinecone
Technology picks that stay maintainable two years out
We decide the stack per engagement, not per trend. Each pick is justified by operating window, team fit and the maintenance contract a senior engineer will have to sign for it.
01, STACK LADDER
Seven layers. Five shortlists per layer.
L2 Application & Web Product UI and admin surfaces
Next.js Remix Astro SvelteKit React Vue TypeScript
L3 Backend & Services APIs, workers, realtime, data path
Node.js Go Python Rust GraphQL tRPC PostgreSQL Redis ClickHouse
L4 Mobile Native and cross-platform
Swift / SwiftUI Kotlin / Compose React Native Flutter Expo EAS
L5 Cloud & DevOps Infra, CI/CD, reliability
AWS Google Cloud Azure Hetzner Terraform Docker Kubernetes GitHub Actions Sentry
L6 Data & Analytics Event pipelines, warehouse, BI
BigQuery Snowflake dbt Airflow Posthog GA4 Metabase
L7 Blockchain & IoT Selective, per use case
Solidity Cosmos SDK Hyperledger MQTT AWS IoT
02, HOW WE PICK
Four filters every pick passes before it reaches code.
Filter 01Operating windowThe pick must make sense across the next two years of maintenance, not just the first sprint.
Filter 02Team fitSenior engineers own the pick. A framework only ships where a named engineer commits to the review contract.
Filter 03Cost honestyWe price infra, inference and operational overhead before the pick is signed off. Surprises go in the scoping document.
Filter 04Exit pathEvery pick is paired with a documented exit, what replaces it if the vendor, model or runtime changes.
CROSS-BENCH
We keep a live bench, not a nostalgic stack.
A small senior group rotates through a bench project every quarter. New providers get measured against the incumbent on the tasks our clients care about. No bench pass, no production.
Current cycleQ2 · 2026
Benchmarked7 LLM providers · 3 vector DBs
Eval tasks42 client-derived
OutcomeAdded 2 picks · retired 1
Stack review
Want to see the right stackfor your workflow?
Share a task. We can return a stack recommendation with a live benchmark within a working week.