Build a language model from scratch. Deploy it as an agent.
A free, self-paced eight-week programme for engineers who want to go from writing their first autograd engine to shipping a deployed, aligned small language model running as an agent. No prerequisites beyond Python.
Eight weeks. Eight things you ship.
Every week ends with something you built, and each deliverable feeds the next: your tokeniser feeds your training run, your model feeds your agent, your agent becomes the capstone.
Your first neural engine
What a gradient actually is, mixed precision and checkpointing: the real fundamentals.
Ships: A working autograd engine and BF16 training loop, written by you.
Build GPT-2 from scratch
Tokenisation, positional encodings, self-attention and the full transformer block.
Ships: A GPT-2 that produces correct logits, verified against the reference.
Make it modern
RoPE, RMSNorm, SwiGLU, grouped-query attention and mixture of experts, plus how vision plugs in.
Ships: A GPT-2 upgraded with RoPE, SwiGLU, GQA and a working MoE layer.
Feed it good data
Train a tokeniser, filter noisy web text, deduplicate at scale and generate synthetic data.
Ships: A custom BPE tokeniser and a filtered, deduplicated pretraining corpus.
Train, scale and align
Scaling laws, multi-GPU training and experiment tracking, then instruction tuning and preference learning.
Ships: A compute-optimal training run and a checkpoint tuned with SFT and DPO.
Connect it to the world
The Model Context Protocol: build the server, define tools and resources, wire your model in.
Ships: A working MCP server exposing tools and data, connected to your SLM.
Give it agency
ReAct-style loops, planning and memory, and when to use one agent rather than many.
Ships: A single tool-using agent that reasons, acts and checks its own work.
Ship it
Orchestration patterns, supervisor and worker routing, and failure handling.
Ships: A deployed multi-agent orchestrator running your own SLM. Live demo day.
Optional research month.
After the core eight weeks, choose one track to go deep on. All three connect to our own research roadmap.
Autoresearch
Multi-modal research agents for regulated data: radiology, genomics, clinical records and scientific literature, with the compliance architecture each domain needs.
World Models
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
Vision encoders, contrastive training, projector-based fusion and multimodal instruction tuning, then deploying a VLM grounded in images.
Start today. Everything here is free.
Is this for me?
Do I need a deep learning background?
No. Comfort with Python is all we assume, and the theory is built from scratch.
How much time does each week take?
Around 8 to 12 hours: pre-reading, the core lab, and finishing your deliverable.
Do I need a GPU?
Weeks 1 to 4 run on free Colab or Kaggle tiers. From week 5 you will want paid GPU time.
What do I walk away with?
An aligned SLM you trained yourself, inside an MCP-connected agent orchestrator you built.
Have a complex AI problem worth solving?
Every engagement starts with a scoping session at no cost. We reply within one business day.