The infrastructure operator's path to practical AI.
You know how to build systems. Now build intelligent ones.
Kindling is a hands-on learning lab for infrastructure professionals who want to understand AI — not academically, but operationally. If you can deploy a service, configure a network, and read a log file, you already have the mental models you need. You just need the right translation.
This project grew from an 84-day partnership between a senior infrastructure architect and an AI coding assistant. Starting from "help me delete duplicate files," that journey produced 98 projects, 8.7 million indexed vectors, fine-tuned models, and AI assistants deployed to field hardware. The breakthrough insight: AI is an infrastructure problem, and the concepts map directly to things you already know.
Kindling makes that journey reproducible.
Kindling is organized as missions — each a standalone, working system that you build and run locally. Every mission produces a tangible artifact: a running service, a searchable knowledge base, a voice interface, an agent that takes actions. Not notebooks. Not theory. Running systems.
| Mission | You Build | Time | Infra Analogy |
|---|---|---|---|
| 00 — Ignition | Chat interface + local model | 15 min | Provisioning your first server |
| 01 — Memory | Embedding pipeline + vector search | 30 min | Building a search index |
| 02 — Retrieval | RAG pipeline over your documents | 45 min | Deploying a query service |
| 03 — Voice | Speech input/output for your AI | 30 min | Adding a management console |
| 04 — Specialization | Domain-specific AI expert | 1-2 hr | Flashing custom firmware |
| 05 — Agents | Tool-calling AI with live actions | 1 hr | Building a control plane |
Run them in order for the guided path, or jump to what interests you.
- Docker + Docker Compose v2 (the
docker composeplugin, not legacydocker-compose) - 16 GB RAM (8 GB minimum with smaller models)
- 10 GB free disk (models + containers)
- GPU optional — NVIDIA for acceleration, but everything works on CPU
- macOS, Linux, or Windows (WSL2)
- For verify scripts:
curlandpython3on the host
git clone https://github.com/enema-combatant/kindling.git
cd kindling
cp .env.example .envEdit .env to select a provider. The default is Ollama (free, local, no account needed):
KINDLING_PROVIDER=ollama
KINDLING_MODEL=llama3.2:3b
KINDLING_EMBED_MODEL=nomic-embed-textSee Provider Guides for other options (OpenAI, Anthropic, Groq).
cd missions/00-ignition
docker compose upOpen http://localhost:5000 and start talking to your AI.
./verify.shEvery mission includes a smoke test that proves the system is functional.
Kindling is provider-agnostic. Switch providers by changing one environment variable — no code changes.
| Provider | Cost | Latency | Setup |
|---|---|---|---|
| Ollama | Free | Local (GPU fast, CPU slower) | Download + ollama pull |
| Groq | Free tier | Fast (cloud) | API key |
| OpenAI | Pay-per-token | Fast (cloud) | API key |
| Anthropic | Pay-per-token | Fast (cloud) | API key |
Each mission references concept documents that translate AI ideas into infrastructure language:
- Embeddings as Hashing — Content-addressable hashes that preserve semantic similarity
- Vectors as Indexes — B-trees to HNSW; query planning to approximate nearest neighbor
- RAG as Query Pipeline —
SELECTwith a semanticWHEREclause - Tokens as Packets — MTU, fragmentation, context windows as buffer sizes
- Prompts as Configs — System prompts are service configuration files
- Agents as Control Planes — The Kubernetes reconciliation loop: observe, decide, act, observe
These aren't dumbed down. They're translated — from one technical domain to another.
Read JOURNEY.md — the narrative of how an infrastructure architect went from zero AI experience to deploying field-ready AI assistants in 84 days, partnering with an AI coding assistant. It covers the mental model shifts, the mistakes that taught the most, and the meta-skill of working with AI effectively.
After completing missions, extensions take you further:
- Fine-Tuning — Train a model on your own data (QLoRA on cloud GPU)
- Security Hardening — TLS, authentication, container isolation
- Monitoring — Prometheus metrics, health checks, Grafana dashboards
- Multi-Node — Distribute inference across machines
- Field Deployment — Take it offline: laptop, phone, constrained hardware
- Mesh Sync — Offline synchronization via LoRa and mesh networks
- No magic. Every mission shows the actual HTTP calls, the actual vector math, the actual prompt construction. When you understand the primitives, you can choose any framework later.
- No notebooks. Infrastructure people think in services, configs, and logs — not cells.
- No framework tax. No LangChain, no React, no build toolchains. Vanilla Python, vanilla HTML, readable by anyone.
- No vendor lock-in. Switch providers with one env var. Run everything locally if you want.
- No "just trust me." Every mission links to concept docs that explain why it works, in language you already speak.
MIT — Use it, modify it, share it.
This project was built in partnership with Claude Code by Anthropic. The entire journey — from first prompt to this repository — demonstrates what's possible when infrastructure expertise meets AI assistance. The best way to learn AI is to build things with it.