Build a language model from scratch. Deploy it as an agent.
A free, self-paced 8-week program for engineers who want to go from writing their first autograd engine to shipping a deployed, aligned small language model operating as an agent. No prerequisites beyond Python.
Not a model you API-called. One you built.
Most AI education teaches you to use models. This program teaches you to build them — from the autograd engine up through the full transformer, then training, aligning, and wrapping the result in an agent system that runs in the real world.
By Week 8, you'll have a working, deployed multi-agent orchestrator — connected to external tools via MCP, running your own SLM as one of its agents.
Eight weeks. Eight things you ship.
Every week ends with something you built. The deliverables accumulate — your tokenizer feeds your training run, your model feeds your agent, your agent becomes your capstone.
Your first neural engine
You'll understand what a gradient actually is, not just how to call .backward(). Mixed precision, checkpointing — the real fundamentals.
Build GPT-2 from scratch
Tokenization, positional encodings, self-attention, the full transformer block. You'll never treat a model as a black box again.
Make it modern
Every upgrade used in today's frontier models — RoPE, RMSNorm, SwiGLU, grouped-query attention, mixture of experts. Plus: how vision plugs into the same architecture.
Feed it good data
A model is only as good as its data. You'll train a tokenizer, filter noisy web text, deduplicate at scale, and generate synthetic data.
Train, scale and align
Scaling laws, multi-GPU training, experiment tracking — then making the model actually useful via instruction tuning and preference learning.
Connect it to the world
The Model Context Protocol is how models talk to external systems. You'll build the server, define tools and resources, and wire your model to it.
Give it agency
ReAct-style loops, planning, memory, when to use one agent vs. many. Your model starts making decisions.
Ship it
Orchestration patterns, supervisor/worker routing, failure handling — then putting it all together into a system that runs in the real world.
You don't need to wait to begin.
These are the materials, tools and compute options we recommend before Week 1. Everything here is free.
- Neural Networks: Zero to Hero — Andrej Karpathy ↗
The best on-ramp to backprop, transformers and language models that exists.
- The Annotated Transformer — Harvard NLP ↗
Attention is All You Need, with every line of the paper annotated in code.
- Dive Into Deep Learning ↗
Free, interactive textbook. Works in PyTorch, JAX or MXNet.
- Chinchilla Scaling Laws (Hoffmann et al., 2022) ↗
How to think about the right model size for a compute budget.
- LLaMA 3 Technical Report — Meta ↗
A well-documented open-weight model. Good case study for Week 5.
- Model Context Protocol — Anthropic ↗
The official MCP spec and quickstart. Essential reading before Week 6.
- Google Colab (free tier) ↗
Enough for Weeks 1–4. T4 GPU, 15 GB RAM.
- Kaggle Notebooks ↗
30 hrs/week of free GPU. Good alternative to Colab.
- Lambda Cloud ↗
Affordable A100/H100 rental for Weeks 5+ when multi-GPU is needed.
Where do you want to go next?
After the core 8 weeks, choose one track to go deep on for a focused fourth month. All three connect directly to Kalman AI's own research roadmap.
Autoresearch
Multi-modal research agents for regulated data.
Build agents that work across radiology, genomics, clinical records, and scientific literature — with the compliance architecture each domain requires.
World Models
Prediction, planning and active inference.
Train a predictive world model, implement active inference, and build a planning loop on top. The principled complement to LLM-heavy stacks.
Vision-Language Models
Extend your SLM into vision.
Vision encoders, CLIP-style contrastive training, projector-based fusion, multimodal instruction tuning — then deploying a VLM that answers questions grounded in images.
Is this for me?
The program is for engineers who want to go beyond using LLMs to actually building, training, aligning, and deploying them. Comfort with Python is assumed. Everything else we build from scratch.
Do I need a deep learning background?
No. Comfort with Python is all we assume. The program builds the theory from scratch — you'll understand why things work, not just how to call them.
How much time does each week take?
Around 8–12 hours: a pre-reading session, the core lab, and cleaning up your deliverable. It's self-paced, so stretch it across two weeks if needed.
Do I need a GPU?
Weeks 1–4 run fine on free Colab or Kaggle tiers. From Week 5 (multi-GPU training) onward, you'll want paid GPU time — Lambda Cloud is the most cost-effective option.
Can I skip weeks?
Each week's deliverable is the input to the next — your tokenizer feeds your training run, your model feeds your agent, your agent becomes your capstone. Skipping is possible, but you'll feel the gaps.
Is there any support?
A live Q&A runs every two weeks, open to everyone regardless of where they are in the sequence. Bring questions on any week, your labs, or your capstone build.
Is the research month compulsory?
No. Weeks 1–8 are a complete, self-contained program ending in a deployed capstone. The research month is for anyone who wants to go further in one specific direction.
What do I walk away with?
A working, aligned SLM you trained yourself, wrapped in an MCP-connected agent orchestrator you built — plus, if you take the research month, a focused project in your chosen track.
Express interest or just start reading.
The program is free and self-paced. Drop us a line to be notified when the next cohort opens, or scroll back up and start with the resources today.