Naptera · Live bootcamp · Fri–Sun evenings
A live bootcamp in two parts. In Part 1 you build a voice AI agent on your laptop. In Part 2 you deploy and operate it on AWS or Azure, with RAG, secure business tools, a knowledge graph, and LLMOps.
Two 7-week partsFri–Sun · 6–8 PM Central
Weekly feedbackDemo and review
Build togetherShow working software each weekThe bootcamp
Part 1 teaches the foundations of AI engineering on your own laptop. Part 2 takes the same project to production in the cloud. Take them back to back, or start at Part 2 if you already ship Python services.
Build a voice AI agent on your laptop: a streaming LLM service, a voice pipeline, search with sources, tools, evaluation, safety tests, and a human handoff.
For you if you can read basic Python and use a terminal. No cloud account needed.
Deploy and run the agent for real: Terraform on AWS or Azure, access control across companies, order and recall tools with a knowledge graph, tracing, monitoring, and recovery.
For you if you finished Part 1, or you pass a short entry check. See Part 2 week by week.
The first class starts here
In Week 1, students start with a realistic wholesale order-support case: a buyer needs to know whether damaged products can be replaced and whether a recall affects their shipment. The application must check the right records, explain its evidence, and know when to involve staff.
“Two cases in order ORD-1042 arrived damaged. Can I get replacements, and was the lot recalled?”
The outcome: a sourced answer, a clear “I don’t know,” or a safe handoff to staff.
What you build in Part 2
Start with a real business case and grow the same application each week. By the final class, you can demonstrate the code, cloud deployment, evaluation evidence, and operating guide.
Ask a question in the browser and watch the model answer appear. Handle slow calls, empty input, errors, and usage.
Turn approved documents into versioned passages. Compare keyword, vector, and hybrid search in Weaviate.
Route a question to retrieval, answer from permitted evidence, show a citation, and abstain when sources cannot support a claim.
Let a buyer ask aloud and hear a grounded answer while keeping the existing text and retrieval paths working.
Check only the buyer’s orders in PostgreSQL and trace recalled supplier lots through a knowledge graph.
Release tested images, inspect traces and alerts, recover from backup, and demonstrate the stack on AWS or Azure.
How it works
Students build an explicit decision path around the model. That makes answers inspectable and lets each data source do the job it is suited for.
Classify the question. Is it about policy, a specific order, a recall relationship, or a request for a person?
Fetch trusted evidence. Use approved documents, an authorized order lookup, or a graph path. A question may need more than one tool.
Compose and check the answer. Name the source, keep customer data scoped, and offer a handoff when evidence or permission is missing.
The application students assemble
What you will learn
Each skill is practised on the same order-support application, so you see how retrieval, tools, safety checks, and operations depend on each other.
Who it’s for
You already write Python. What you want is the engineering around the model: where its facts come from, what it is allowed to do, and how you know it still works next week.
You build web or backend software and have been asked to add an assistant. You want it to answer from your company’s data and stay inside its permissions.
You can prototype in a notebook. You want to package, test, deploy, and monitor the same idea as a service other people rely on.
You are scoping an AI product for your own company or a client. You want to know what a dependable version costs to build and run.
Newer to code? Start with Part 1: you need to read basic Python and use a terminal. Looking for a no-code course? This is not it. Talk with us first and we will tell you honestly.
Part 2 curriculum
Friday introduces the problem and the engineering decisions. Saturday is a guided build. Sunday is for finishing the checkpoint, running the tests, and showing it for feedback. Students complete the assignment in their own repository before the next Friday.
Trace the buyer’s damaged-order and recall question, map where private data lives, and make the first application boot from a clean clone.
Friday Map the buyer journey, system boundaries, failure cases, and AWS or Azure path.
Saturday Build the Python API and browser page; run Docker Compose, a health test, CI, and a Terraform plan.
A reviewer can boot the stack, see the passing GitHub Actions check, and inspect the infrastructure plan.
Learn tokens, inference, context, prompting, latency, usage, and why a model can sound certain without knowing a fact.
Friday Compare model responses with and without evidence; define when the app should answer or abstain.
Saturday Stream a server-side model response, test timeout behavior, publish a versioned image, and deploy it.
Show a streamed answer, a graceful failure, a passing CI run, and a rollback to an earlier image.
Learn how documents become chunks and vectors, and why a close match is a candidate source rather than proof.
Friday Compare keyword, vector, and hybrid retrieval; define expected sources for ten test questions.
Saturday Ingest versioned documents, build a search view, test access filters, and update the cloud stack.
Find an exact passage and its version, then rebuild a deleted search index from the source documents.
Connect the reviewed Week 3 index to an agent decision path. A spoken question can retrieve policy and produce a short sourced answer.
Friday Add a browser audio room, short-lived access token, agent participant, transcription, and speech output.
Saturday Add the policy search tool, source cards, refusal and handoff logic, and a golden question set.
Demonstrate a cited policy answer and an unsupported question that produces an honest abstention.
Combine live order facts with approved policy while keeping one organization’s records invisible to another.
Friday Create Postgres tables and migrations; bind the signed-in buyer to read-only order lookup.
Saturday Route order and policy questions, draft a support case after confirmation, and hand uncertain cases to staff.
Show a valid order answer, a denied cross-organization lookup, an evaluation result, and a human handoff.
When a supplier recalls a lot, identify affected customer orders and show the exact shipment path. Learn to model relationships, query the graph, check data freshness, and protect each buyer’s records.
Friday Model supplier, lot, shipment, product, and order relationships in a knowledge graph.
Saturday Add a permission-checked recall tool and combine graph facts with approved recall policy.
Trace a seeded lot to affected orders and explain what happens if the graph is stale or unavailable.
Find why an answer is slow or wrong, then demonstrate that another engineer can operate the complete system.
Friday Follow one request through transcription, routing, retrieval, databases, model response, and speech.
Saturday Deploy pinned images, trigger an alert, restore a backup, rebuild derived data, and rerun the tests.
Show the live app, a trace, a dashboard, a safe failure path, and a completed restore runbook.
Final demonstration
Students finish with a repository, cloud deployment, evaluation report, and operating guide. The final review follows one buyer request through the whole stack.
Dates and times
Three 2-hour evening sessions every week, Friday to Sunday, so you can keep your day job and plan your calendar before you enrol.
Tuition
Payment agreement available if needed.
Ask about the curriculum, request a call, or discuss a payment agreement.
Share a few details and we’ll contact you within a day to talk through the class and next steps.
After the class
The final project gives you working software to show. If you want help with your next step, choose the path that fits your goal.
Live class format
Every week combines an engineering concept with an application students can demonstrate. Plan for about eight to ten hours a week, including live sessions and independent work.
Review the user problem, draw the data flow, compare approaches, and decide what safe behavior looks like.
Implement the week’s feature, run the stack, test a failure case, and see the expected checkpoint.
Get help finishing the checkpoint, run the tests, and show a short demo for feedback. Wrap up the assignment in your own repository before Friday.
Before you start
You should be comfortable with basic Python, Git, and a terminal. You do not need to have built a production AI application before the class.
For Part 1, you need to read basic Python and use a terminal. For Part 2, you should be comfortable with Python, Git, and Docker. You do not need machine learning experience or to have shipped an AI application. If you are unsure, send us a message and we will look at it with you.
No. If you already build Python services with Git and Docker, you can start at Part 2 after a short entry check. If you are newer to code, start with Part 1.
About 8 to 10 hours: three live sessions of 2 hours, Friday to Sunday from 6:00 to 8:00 PM Central, plus 2 to 4 hours to finish the assignment. Weeks 2, 5, 6 and 7 include cloud deployment and run toward the top of that range.
Tell us before you enrol if you expect to miss sessions, and we will tell you how this cohort handles it.
A GitHub repository, an application deployed on AWS or Azure, an evaluation report, a monitoring dashboard, and an operating guide, plus a recorded final demonstration.
Model calls and a small cloud host cost money on your own accounts. We show the cost of one request in Week 2, set a spending alert before you deploy, and remove resources at the end.
You choose AWS or Azure and deploy to that one provider for the whole course.
A payment agreement is available if you need one; ask us before you enrol. Ask us for the refund terms before you pay.
Ask us before you enrol and we will tell you what is included.
