AI Platform Masterclass · prototype
Engineer the surroundings, not the model.
The model is fixed. Everything that decides whether it ships is around it. Each lesson is an interactive you operate first, then we map the concept to the tools your stack already runs: LangChain, LangGraph, Langfuse.
Engineer the surroundings
The model is fixed. Everything that decides whether it succeeds is around it: the prompt, the context, the loop, and the harness.
Prompt engineering
Assemble a prompt from parts and watch each one buy back reliability. It is not word tricks, it is reliability engineering.
The uncomfortable truth
A clever prompt that parses correctly 88% of the time is a broken feature.
Toggle prompt parts on and off, watch the failure rate move.
Context engineering
The window is a budget, not a bucket. Learn what to keep, and watch a naive truncation evict the actual question.
The uncomfortable truth
Most "the model ignored my instructions" bugs are really "your instructions got evicted".
Spend a fixed token budget, then watch what falls out when you overflow it.
Loop engineering
An agent is a loop: think, act, observe, repeat. Learn the two ways it fails and the four things that stop it.
The uncomfortable truth
A loop with no stop condition and no progress check is a bill with no ceiling.
Step the loop, inject a tool error, and flip reflection on to watch it converge instead of spin.
Harness engineering
The model is the engine. The harness is the car. Same model, different harness, and the success rate goes from 8% to 90%.
The uncomfortable truth
Most "the AI cannot do this" is not a model problem. It is a harness problem.
Bolt capabilities onto a bare model and watch a task it kept failing start to succeed.
Secure it for a bank
The model cannot tell your instructions from the data it reads. Everything an attacker does starts there.
Prompt injection
The number-one AI security risk, in three tiers. See it work, feel why no prompt-level defense is complete, then contain the blast radius by design.
The uncomfortable truth
You cannot prompt your way to safety. The model obeys the ticket, and the ticket is the attack.
Inject a support ticket, watch the agent refund $5,000, then stack defenses and see which never close.
Guardrails & policy
The deterministic code around the model that decides what reaches a user. Set the threshold, feel both failure modes, then move the rules into a policy engine.
The uncomfortable truth
A guardrail that blocks everything is theater. One that blocks nothing is decoration. The craft is the threshold.
Slide a jailbreak threshold and watch it block real customers or miss real attacks. There is no free setting.
MCP & tool auth
A tool call is an API call, and the model is the least trustworthy caller you will ever have. Whose identity does it carry, and what can a leaked token do?
The uncomfortable truth
Authenticating the agent is not authorizing the call. Most agent tools run as a god account nobody scoped.
Make the model read a stranger’s balance, over-grant a token until it is a skeleton key, then leak it.
Operate it in production
Once it runs for real, you have to see inside it. This is where a demo becomes a platform.
Why concept-first
Your company runs LangChain, LangGraph and Langfuse today. Tools change. The reason every lesson teaches the idea before the tool is so that when the stack is swapped, your knowledge still applies. Learn the middle column, and the left and right are just syntax.