Your own model
An open 27B model fine-tuned on your domain, with a registry card: weights, metrics, licence, version.
Five things you walk away with
You build the first three by hand. Microsoft verifies the fourth. The fifth starts earning.
A production app
A chat over your documents with retrieval and citations. Deployed to the cloud, reachable by link, running on your model.
An eval suite
A test suite that catches hallucinations and regressions before release and blocks the deploy when quality drops.
Certification
Microsoft AI-103. The badge is published on LinkedIn and verifiable by any employer.
A marketplace listing
Your own offer in the Microsoft catalogue, where enterprise buyers look for contractors.
Eight weeks
A working application exists by the end of week one — before any theory. From there you take it apart layer by layer and replace each layer with your own.
- Week 1Your first cloud applicationPython, git, Docker, access. Deploy the chat template, load your documents, get a live link
- Week 2How document retrieval worksChunking, vector and hybrid search, citations, the prompt layer
- Week 3Cost and latencyTokens and delays as a design constraint. Caching, choosing the right model per task
- Week 4Fine-tuningQLoRA on a 27B model: data collection, the run, the usual ways to break it and how to catch them
- Week 5Your own endpointServe the model on vLLM, wrap it in a compatible API, swap it in for the vendor's
- Week 6Quality gatesBuild an eval suite and make it a release gate. This is the first thing interviewers ask about
- Week 7Agents and toolsMulti-step scenarios, calling external systems, Model Context Protocol, failure recovery
- Week 8Project defencePresent the system to a technical reviewer and explain its limits to a non-technical executive
Three tracks run in parallel
Engineering is the main one. The other two take half an hour a week and run in the background, so everything lands at the same time.
Engineer
Eight weeks of project work: from a stock template to your own model in production with measured quality.
AI-103 certification
An AI tutor inside your workspace walks you through the exam domains, remembers your mistakes and returns to weak spots. The next domain stays locked until the previous one is cleared.
Microsoft partnership
Entity verification takes weeks, so the application goes in from week two. By graduation your offer is live in the catalogue.
An adaptive tutor, not a video library
Every student gets a personal AI tutor running on the same 27B model you fine-tune. It works from what you actually did, not from what a syllabus assumes you did.
Runs on your model
The tutor is served from the shared 27B endpoint inside your workspace. It reads your code and your logs, not a generic textbook.
Remembers your mistakes
Every wrong answer and every failed run is kept. The tutor returns to those weak spots until they hold, then moves on.
Mastery gates, not deadlines
AI-103 domains open one at a time. The next stays locked until the previous is cleared — the way the exam itself is scored.
Checks by running code
When you claim something works, the tutor executes it and reads the real output. A plausible answer is not accepted; a correct one is.
The same tutor grades your weekly project checkpoints, so exam preparation and engineering work feed one record.
Real hardware, not a simulator
Fine-tuning cannot be learned in theory. Every student gets a dedicated 40 GB slice of an NVIDIA A100 on the Lambda cluster, a partner of the NVIDIA Inception programme. No queue for a card and no neighbour eating your memory.
A 27-billion-parameter base model is available around the clock through a shared endpoint: your applications and your tutor run on it. Training slots hold QLoRA of the same model — enough for fifteen to thirty full runs per person.
AcceleratorNVIDIA A100 80 GB
Base model27B, always on
Training slot40 GB
Slot hours per student179
Full fine-tuning runs15–30
Graduation run8 × A100, 640 GB
EnvironmentvLLM · Jupyter · Docker
What this work sells for on the Microsoft catalogue
Two live listings from the Microsoft Marketplace. Both cover exactly what you learn in the first five weeks.
RAG Implementation: 4-Wk Proof of Concept
A custom chatbot on Azure OpenAI with advanced RAG, integrated with the client's systems. Four weeks.
Azure AI RAG & Enterprise Search
An enterprise document search MVP. The listing names the stack: Azure OpenAI, Cognitive Search, Bot Framework, Azure Storage. Four weeks.
This is a market reference, not an income guarantee. We take you to the point of publishing your own listing and explain both routes: a listing-only offer where the deal closes directly with the client, and a transactable offer billed through Microsoft. What happens next depends on Microsoft and on you.
Who is hiring right now
The shortage is not in theory but in delivery: there are three times fewer people who can take a model to a working system than there are open roles.
3.2 : 11.6M open positions against 518K qualified candidates
1.3MAI job openings created in two years, per LinkedIn
$350–550Kestimated annual package for a Forward Deployed Engineer at OpenAI
In Microsoft job postings the keyword is Azure AI Foundry, usually alongside Copilot Studio. They are not hiring researchers but engineers who assemble and deploy agents. The new AI-103 exam is built on exactly that.
At OpenAI the largest engineering role is Forward Deployed Engineer. The loop consists of coding on agentic scenarios, system design under cost and latency constraints, and a conversation with a non-technical executive. Week eight of this course is built to that list.
Deploying a model inside the client's own perimeter is valued separately: banks, healthcare and the public sector do not send data to a public API. That cannot be learned on a course that lives on a single cloud key.
Price
One price for eight weeks. Compute, exam preparation and support through to publishing your listing are included.
$1,800
first cohort · $2,400 thereafter
- 179 hours of a 40 GB A100 training slot
- A 27B base model available around the clock
- 8 weeks of instruction in a group of up to 20
- AI-103 preparation with a personal tutor
- A graduation run on an eight-accelerator node
- Code and project review at every stage
Paid separately
AI-103 examdirectly to Microsoft
Your own Azure subscription$30–60 per course
Partner Center registrationper Microsoft rules
Spend on your own Azure subscription stays small because the heavy runs happen on our cluster, not in your cloud. An Azure billing profile is required at intake — week one starts with it.
Who stands behind the programme
Everything below can be checked in a public register. Nothing here rests on our word alone.
Our UK entity is registered on the UK Register of Learning Providers and holds a verified UK Provider Reference Number. Look it up yourself at ukrlp.co.uk. To be precise about what this is: the register confirms a verified legal identity. It is not a quality accreditation and not an endorsement of any course by the UK Department for Education.
Membership in NVIDIA's programme for technology companies. This is where the academy's compute comes from. The programme publishes no member directory, so we send a confirmation letter on request.
Publisher status in the Microsoft commercial marketplace — the same route graduates use to publish their own offers. Confirmation from Partner Center on request.
Ask for any confirmation letter at dev@hybt.ai. We send them to a named recipient, dated.
Questions
Short answers to what applicants ask most. Anything else: write to us.
Do I need my own computer with a GPU?
No. A laptop with a browser is enough. All compute runs on our NVIDIA A100 cluster: students connect to a prepared environment through the browser or over SSH from their own editor.
Which model do students train?
An open model with 27 billion parameters. The base version is available around the clock through a shared endpoint, and QLoRA fine-tuning runs on a dedicated 40 GB slice of an A100. Each student has enough hours for 15 to 30 full runs.
What is included in the price?
179 hours of a 40 GB A100 training slot, round-the-clock access to the 27B base model, eight weeks of instruction in a group of up to 20, Microsoft AI-103 exam preparation, a graduation run on an eight-accelerator node and code review at every stage.
What is paid separately?
The AI-103 exam is paid directly to Microsoft. Your own Azure subscription costs $30–60 across the course because heavy runs happen on our cluster. Partner Center registration follows Microsoft's own rules.
Which certification do graduates take?
Microsoft AI-103, Developing AI Apps and Agents on Azure, leading to the Azure AI Apps and Agents Developer Associate credential. It replaced the retired AI-102 and devotes roughly a third of its weight to the generative and agentic domain.
How do graduates monetise the skill?
By publishing their own offer in the Microsoft Marketplace catalogue. We cover both routes: a listing-only offer where the deal closes directly with the client, and a transactable offer billed through Microsoft.
What you need coming in
We do not require experience with models. We require that you can write code and read logs.
- Python at the level of confidently reading and editing someone else's code
- Command line, git, a working understanding of Docker
- An Azure account with a billing profile attached
- English sufficient to read documentation
- 10–12 hours a week, four of them synchronous
- A laptop; all compute runs on our cluster
Apply for the first cohort
Selection is a 30-minute technical interview: we go through your code and discuss which problem your graduation project will solve.
20 seats remaining · starts 12 October 2026
Field notes
What we measure on our own hardware while building the programme. Numbers shown, method stated.
Memory bandwidth, not memory size, decides which model runs well on a DGX Spark
A DGX Spark holds 128 GB, so a 72B model fits. It still generates about five tokens per second. Here is the arithmetic that explains why, and what to run instead.
Who runs the academy
A transnational group working in applied AI, sovereign edge computing and industrial automation.
HYBT Technologies, Inc.
Newark, Delaware. The group's technology hub and vendor partnership entity. Member of NVIDIA Inception — the source of the academy's compute.
Digital Qazaqstan LTD
London. Intellectual property, research and publishing. Member of the Microsoft AI Cloud Partner Program.
AL SAFI FZE
Umm Al Quwain Free Trade Zone. MENA operations hub, licensed for AI development and AI education. Operator of the academy.
Partner programmes: NVIDIA Inception — access to compute infrastructure and the NVIDIA ecosystem. Microsoft AI Cloud Partner Program — publisher status in the commercial marketplace, through which graduates publish their own offers. RAKTN — regional technology partnership.
These programmes publish no public member directories, so confirmations are issued by letter on request: dev@hybt.ai.