You vibe coded it. It works. Now make it survive success.

• TypeScript
• Python
• Go
• Rust
• Node.js
• Bun
• Ruby on Rails

TypeScript end to end for most products: one language across frontend and API cuts handover cost and hiring risk. Python where the workload is AI-heavy. Go where raw throughput decides the architecture, and Rust on hot paths where every millisecond is money. We run Node.js and adopt Bun where its speed pays. Our engineers shipped and scaled Ruby on Rails at Volopay, and we still support Rails codebases in production.

• React
• Next.js
• Astro
• React Native
• Expo
• Tailwind CSS
• shadcn/ui
• Radix UI

React and Next.js for web products. Astro for content-heavy sites where load speed is the feature. React Native with Expo for iOS and Android from one codebase, native modules where the product demands them. Tailwind CSS with shadcn/ui and Radix for design systems that ship fast. Real-time and streaming interfaces with WebSockets and server-sent events, because AI products that make users wait for a full response feel broken.

• Convex
• Node.js
• Hono
• Postgres
• Drizzle
• Redis
• Zod
• Better Auth
• FastAPI
• Supabase

Two backend tracks, chosen per product. Convex for real-time serverless products: queries, mutations, actions, a scheduler, cron jobs, storage, and live subscriptions. Node.js services with Hono and typed oRPC over Postgres and Drizzle ORM for relational, retrieval-heavy products. Postgres is extended with pgvector for vector search and full-text search, so most AI features ship without a second database. Redis for caching and queues. Zod for end-to-end validation, Better Auth for authentication, and FastAPI or Supabase where they fit the job.

• OpenAI
• Anthropic
• Gemini
• Llama
• Mistral
• DeepSeek
• ElevenLabs

Model choice is an engineering decision, not a loyalty program. OpenAI GPT and Anthropic Claude for frontier reasoning, Google Gemini where multimodal or context length wins. Open-weight models, Llama, Mistral, DeepSeek, self-hosted when privacy or unit economics demand it. Specialist models where the product needs a specialist: ElevenLabs and dialogue-native TTS for production voice, Whisper for speech to text, dedicated embedding and reranking models for retrieval quality, image and vision models for documents and floor plans. Every model sits behind a provider-swappable abstraction, so switching is a configuration decision, not a rewrite.

• Vercel AI SDK
• Qdrant
• LangGraph
• Hugging Face

AI features are built on the Vercel AI SDK and Convex Agent and RAG components: structured outputs, agentic tool use, streaming, and embeddings. Retrieval on pgvector by default, with Cohere reranking and Tavily web-search fallback, and managed vector stores like Qdrant at scale. Orchestration with LangGraph or custom code, tool connectivity over MCP, the protocol now standardising how AI systems talk to software. And the part that decides whether AI survives production: evals before ship, tracing in production, structured outputs, semantic caching, and model routing to control cost.

• Ollama
• FFmpeg

We deploy AI through AWS Bedrock, Google Vertex AI, and Azure OpenAI where enterprise controls matter, direct APIs where speed matters, and vLLM or Ollama where self-hosting wins. Self-hosted GPU inference on RunPod with a scale-to-zero strategy, and image models through fal.ai. Media assembly and normalisation with FFmpeg.

• AWS
• Google Cloud
• Azure
• Docker
• Coolify
• Kubernetes
• Turborepo
• Terraform
• GitHub Actions

AWS first: KUMO is an AWS Partner and runs its own SaaS, CampaignHQ, on AWS in production, with S3, EC2, CloudFront, and Route 53. Also Google Cloud, Azure, and client infrastructure including on-prem. Docker everywhere, Coolify and Kubernetes for deployment, serverless and edge where they fit. Monorepos on Turborepo, infrastructure as code with Terraform, background jobs on Trigger.dev, and delivery through GitHub Actions CI/CD.

• Sentry
• OpenTelemetry
• Playwright

Sentry for application errors, Langfuse for LLM traces, model metadata, and token usage, and OpenTelemetry for distributed tracing. End-to-end and AI-driven browser testing with Playwright. Cost, latency, and error tracking by workflow and model, so problems surface before customers see them.

• Stripe
• Razorpay
• Resend
• Discord

Billing and payments with Stripe and Razorpay, subscriptions, webhooks, tax, and invoicing. Transactional email through Resend, team workflows over Discord, and the Google ecosystem, OAuth, Drive, and Sheets, wired in where products need it.