Guides, prompts, and reading material we use at MTechZilla when building with LLMs, shared publicly so others can borrow from them.
The centrepiece is How LLMs Work - a 17-part written course, in plain language, for engineers who want to build on LLMs with real understanding rather than use them blindly. Read it in order; every part builds on the one before, and each ends with a table of memory hooks for later recall.
From raw text to the token on your screen: what the machine is, how it learns, and what it is doing when it answers you.
| # | Guide | Time | What it answers |
|---|---|---|---|
| 1 | Data Preparation | 2 min | Where 10-15 trillion training tokens come from, and who cleans the data before a smart AI exists to clean it |
| 2 | Tokenization | 4 min | Subwords and BPE, why the tokenizer is frozen forever, why models can't spell "strawberry", and why Hindi prompts cost 2-3× more |
| 3 | Pre-Training | 5 min | The machine (billions of dials), the game (fill-in-the-blank), and the parts - embeddings, attention, MLPs, the stack |
| 4 | Post-Training: SFT, RLHF, Verifiable RL | 4 min | How an autocomplete becomes an assistant, and where "thinking" actually comes from |
| 5 | Inference: The Loop, Temperature, Top-p | 4 min | One token per pass, no re-rolls - and what the two sampling knobs really do |
| 6 | Chat Templates and System Prompts | 2 min | Why instructions have power at all, why a base model ignores them, and how far a system prompt can be trusted |
| 7 | Context Windows | 2 min | The desk vs. the library, why long chats forget, and "lost in the middle" |
| 8 | Embeddings and Vector Databases | 3 min | How semantic search works, the RAG architecture, and which parts are your code's job |
From a working demo to a production system: the disciplines that decide whether it survives real traffic.
| # | Guide | Time | What it answers |
|---|---|---|---|
| 9 | Production RAG | 6 min | Chunking strategies, why retrieval quietly fails, reranking, and how to measure quality instead of guessing |
| 10 | Structured Output and Reliable JSON | 4 min | Why "respond in JSON only" is a prayer, and the ladder of fixes that turns it into a contract |
| 11 | Tool Calling | 4 min | The model never calls anything - it emits a slip your code executes, and where the trust boundary sits |
| 12 | Model Context Protocol (MCP) | 3 min | The N×M integration problem, USB as the analogy, and what MCP does not change |
| 13 | How AI Agents Work | 4 min | The loop given a goal instead of a question, where agents shine, and when you shouldn't build one |
| 14 | Context Engineering | 3 min | What the model should see right now - the four moves, the three tiers, and prompt caching |
| 15 | Evaluating LLM Applications | 3 min | Golden sets, the three graders, LLM-as-judge, and reward hacking in your own pipeline |
| 16 | Prompt Injection and Guardrails | 4 min | Why instructions ARE data, the attack ladder, and why the fix cannot live inside the model |
| 17 | Cost and Latency Optimization | 3 min | Model routing, prompt caching and prefix drift, and the arithmetic behind a client quote |
Roughly an hour end to end. If you are shipping something specific rather than learning the whole pipeline, jump straight to the guide for that problem - each one stands alone, and links back to the earlier guide when it leans on an idea from it.
- prompts/ - production prompts we run ourselves, written to be copied and adapted. Currently: summarising meeting transcripts into decisions, action items, and owners.
- resources/ - the curated reading list the guides were built from: 35 talks, articles, and docs grouped by topic, each with a note on what it covers and how long it takes.
- guides/ - the folder index for the course above.
Issues and pull requests are welcome - a correction to a guide, a better prompt, or a resource worth adding.
MIT - see LICENSE.