Uses
What we actually use to ship AI.
Our defaults — models, frameworks, infra, observability. Updated when we change our minds. Not a complete list of everything we know.
Models
Primary, secondary, and fallbacks we reach for first
Anthropic Claude (Sonnet, Opus, Haiku)
Default for agents, RAG, and anything where reasoning quality matters.
OpenAI GPT-4.1 family
Secondary for variety, vision tasks, structured-output prompts.
Open-source via Together / Groq
Llama 3.x, Qwen 2.5 when cost optimization or on-prem matter.
Embeddings
Cohere embed-v3 and OpenAI text-embedding-3-large as defaults.
Agent + orchestration
What we use when one model call isn't enough
LangGraph
Default for multi-step agent workflows with branching and tool use.
Anthropic Agents SDK
When we want Claude-native tool orchestration without an extra layer.
n8n (self-hosted)
Workflow automation — invoices, leads, content, ops glue.
Vapi / Retell / Twilio
Voice agents — inbound, outbound, IVR replacement.
Retrieval + memory
How we keep agents grounded
pgvector
Default. If you already have Postgres, this is where we start.
Pinecone / Qdrant
When pgvector doesn't scale or hybrid search needs more headroom.
Cohere Rerank v3
Reranking the top-50 to top-5. Cheap, dramatic precision lift.
Redis
Conversation state, session memory, rate-limiting.
Eval + observability
How we know it's still working at 11pm on a Saturday
Langfuse
Default observability. Self-hosted available when needed.
Promptfoo
Local prompt regression tests in CI.
Ragas
Faithfulness, answer relevance, context recall on RAG pipelines.
Sentry
Boring app-level errors. Boring is good.
Infra + dev
What we ship things on
Vercel
Default Next.js hosting. Edge functions where latency matters.
Modal
Python services, batch inference, fine-tuning runs.
Supabase / Neon
Managed Postgres with pgvector. Quick to stand up.
Cloudflare Workers + R2
Edge cases — pun intended — when Vercel isn't the right home.
Docker
Sandboxes for Claude Code and other untrusted agent execution.
Frontend
What ships in the browser
Next.js + React
App Router. Default for everything customer-facing.
Tailwind CSS
v4 with CSS-first theme. Brand tokens as CSS variables.
shadcn/ui primitives
Vendored, not installed. Easier to customize, no surprise updates.
Vercel AI SDK / assistant-ui
Chat UIs when we don't want to rebuild scrollers and streams.
framer-motion
Motion. Sparingly.
Working tools
How the team actually operates
Claude Code + Cursor
Day-to-day code editing. Yes, our agents write some of our agents.
Linear
Issues, sprints, projects. Slack-native for ops.
GitHub
Source of truth. Every client engagement gets its own repo.
Toggl
Time tracking for hourly engagements. Boring is good.
Cal.com
Discovery and technical deep-dive scheduling.
