01The formats

Three ways to build the capability in-house

Every format ends in certification against shipped work and a written capability debrief. We recommend one against your baseline in the scoping call.

Cohort waves Core programme

AI engineer conversion cohort

Selected builders, typically your top 10–20%, progress part-time around delivery: self-paced study, live instructor checkpoints, exercises against your approved use cases, and a capstone that ships: a deployed, evaluated LLM system.

ScaleWaves of 10 to 40; parallel streams at scale CoversModel behaviour, RAG, evaluation, agents, deployment Leaves withShipped capstones, certification, capability debrief

Per cohort · quoted after scoping Scope a cohort →

8–12 weeks For your leads

AI leader residency

One to six future AI leaders work with our senior architects: weekly coaching, applied labs, architecture reviews, and a capstone from a real internal workflow. Each resident turns one workflow into a reference build later cohorts train on.

Scale1 to 6 senior engineers per wave CoversWorkflow qualification, baselines, adoption, the demo-to-production gap Leaves withReference builds, internal instructors, your standards

Per resident · priced as senior design work Nominate residents →

Per seat

Enterprise Academy licence

The full Academy bundle: Full Stack AI Engineering (92 lessons), Agentic AI Engineering (5.0-rated), the Agentic Development course, and AI for Work, with live kickoff, monthly mentorship clinics, and leadership analytics. Lifetime updates as our client work evolves.

ScaleHundreds to thousands of seats CoversThe complete curriculum, self-paced with role paths Leaves withOne shared foundation and progress analytics

Per seat · volume bands Discuss seats →

At 1,000+ developers: residency seed → 20–40-seat pilot → waves of 50–250 in parallel streams → top graduates co-instruct as internal champions. Per-developer cost drops as the programme scales. A Python foundations route exists for mixed-skill cohorts.

02The system

Two steps: use AI to build software, then build AI into it

Step 1 makes every developer effective with coding agents, and its debrief names who belongs here. You are on the Step 2 page.

	Step 1 · Agentic developers	Step 2 · AI engineers ← this page
Who	Every software developer, no AI background needed	Selected builders, typically the top 10–20%
They learn	To plan, direct, review, and verify coding-agent work under one standard	To design, evaluate, ship, and maintain production LLM systems
Outcome	Ship faster on today's roadmap	The capability you cannot hire: AI products and systems, built in-house
Format	1-day or 2–3-day bootcamp · per-seat licence	Part-time cohort waves · 8–12-week residency · per-seat licence
When	Now: it pays back on work already in flight	After Step 1, or now if your devs already work seriously with agents

Not there yet? Start with Step 1 → Same instructors, same standard, one scoping call covers both.

03Why convert

Deployed AI pays back. Pilots without owners don't.

The pilots graveyard

MIT's 2025 study attributes the failure of ~95% of enterprise GenAI pilots to a “learning gap”: workflow and people integration, not model quality. S&P Global finds 42% of enterprises abandoned most AI initiatives in 2025, up from 17%. Pilots stall because nobody owns production: design, evaluation, deployment, operations.

MIT NANDA 2025 · S&P Global 2025

Deployment pays

McKinsey finds only 7% of firms have deployed AI organisation-wide while 65% are stuck at proof of concept, yet 95% of deployed AI meets or exceeds ROI expectations, and 94% of firms are increasing spend into exactly that gap. The missing input isn't budget or models. It's people who can carry systems to production.

McKinsey 2026

You can't hire your way out

AI engineer is the #1 fastest-growing US job, postings up 143% year on year, with a 56% wage premium for AI skills, and LinkedIn's own data shows the most common feeder role is the software engineer. Your conversion pipeline is already on payroll; a new hire spends a year learning what your team already knows.

LinkedIn 2026 · PwC 2025

Capability compounds

The first trained cohort sets the engineering standard every later cohort inherits. Graduates seed later waves as internal champions and co-instructors, per-developer cost drops as the programme scales, and the capability stays in-house, with no dependence on us for the next use case.

Programme design · scale path above

04Straight answers

The questions your board will ask

“Why not just hire AI engineers?”

The role is barely three years old, supply is thin, the labs hire each other first, and frontier comp is extreme. Meanwhile your developers already hold the domain, systems, and codebase knowledge AI systems run on. Conversion wins on cost, speed, retention, and context.

“Most AI initiatives fail anyway.”

They fail for skills reasons: MIT attributes the 95% pilot-failure rate to a learning gap, not model quality, while 95% of deployed AI meets or exceeds ROI expectations (McKinsey). Capability-building is precisely the variable that moves you across that divide.

“Training doesn't stick.”

Lecture-style training doesn't. This programme certifies against shipped capstones (a deployed, evaluated LLM system per engineer) with exercises drawn from your approved roadmap, and closes with a written capability debrief. Nobody is certified for attending.

“We can't take engineers off delivery.”

Conversion runs part-time around delivery: self-paced study plus scheduled live checkpoints, shaped with engineering leadership. The residency is a defined weekly commitment for one to six seniors, not weeks out of the roadmap.

“How do we choose who to convert?”

Run Step 1 first: its capability debrief names the builders ready for this programme. Otherwise we baseline and segment the population in scoping (developers, data scientists, existing agent users) and agree the first cohort with engineering leadership.

“Generic courses won't fit our stack.”

Depth and emphasis are agreed against your engineers' baseline; electives are added where your stack needs them; exercises come from your roadmap, never generic demos. Every pattern we teach appears in a system we built; see below.

05Full transparency

The curriculum, module by module

Depth and emphasis are agreed against your engineers' baseline in scoping. Exercises come from your approved roadmap, never generic demos.

01Foundations and model behaviourcore

1Models, prompting, and the limitations that shape engineering decisions
2Security, prompt injection, and responsible technical use
3Choosing the right technique for the task: prompting, RAG, fine-tuning, or agents

02Systems: retrieval and application patternsbuild

1Retrieval-augmented generation, from first pipeline to production patterns
2Framework and architecture trade-offs; context and application design
3Application-oriented exercises against approved use cases

03Evaluation and reliabilityrigour

1Test-set design and evaluation as an engineering discipline
2Generation and retrieval evaluation; LLM-as-judge and regression thinking
3Observability, so output is never accepted on intuition alone

04Agent engineeringadvanced

1Agentic patterns and when agents beat simpler techniques; tool use and function calling
2Orchestration and multi-agent systems; MCP integrations; memory and state design
3Agent evaluation, guardrails, and human-review boundaries

05Optimisation and deploymentproduction

1Fine-tuning and optimisation; deployment and cost control
2Monitoring in production, and staying current as the field moves
3Electives where your stack needs them: large codebases, domain integrations, compliance

06Scoped to your baselinecustom

1Depth and emphasis agreed against the engineers' baseline during scoping
2Module weighting follows your use cases, not a fixed timetable
3No stage is filler for the next

07Capstone: certified against shipped workcertified

1Every module lands in capstone work: a deployed, evaluated LLM system
2Certification against shipped capstones, not attendance
3A written capability debrief closes each cohort

Outcome: engineers who design, evaluate, ship, and maintain production LLM systems, certified against a deployed capstone.

06We teach what we ship

Every pattern we teach appears in a system we built

21+ AI applications shipped to production across finance, government, law enforcement, healthcare, and industry. The curriculum is drawn from that work, which is why your training partner must be a practising builder.

Banker-reasoning engineBanking · Europe

Agentic financial-intelligence platform routing 1M+ words per company across four frontier model families; ~800 prompts and 20-step workflows per analysis.

Teaches: routing, orchestration, and evaluation, from idea to live demos with global banks.

Dialog InsightMartech · Canada

Multi-month production-agent enablement: weekly advisory, 16+ hours of guided code transfer, six runnable reference repositories, a 25-scenario benchmark.

Outcome: working code the team owns: a production-shaped agent stack, runnable and auditable.

Document extraction at scaleMortgage · Canada

50+ bilingual document types for underwriting; three independent extraction passes with consensus; never-overwrite Salesforce integration; data-residency compliant.

Teaches: structured outputs, consensus architectures, human-review boundaries.

“The most comprehensive textbook to date on building LLM applications; every essential topic in an AI engineer's toolkit.”

Jerry Liu

Founder & CEO, LlamaIndex, on our book

“An indispensable guide for anyone venturing into the world of LLMs, covering everything from theory to practical deployment.”

Senior Data Engineer

We ran this conversion line on ourselves first

Our delivery arm needed more production AI engineers than the market could supply, so we built a conversion line: Learner → Author → Instructor → Engineer. 15+ of our own engineers came off it. Work is the audition, and your cohorts are held to the same standard.

Louie Peters Co-CEO · Instructor

Co-author, Building LLMs for Production (10k+ copies). Chief AI Officer for a banking agentic-engine build. Ex-J.P. Morgan VP, Research.

Louis-François Bouchard CTO · Instructor

Co-author of the book; 210k followers; ex-PhD at MILA (Yoshua Bengio's lab). Built our 100k+ member community.

Fabio Chiusano Senior AI Engineer

NLP lead; ex-founder & Head of Data Science. Joined as a contributing author; proof the pipeline works.

Assel Kashkenbayeva Deployment lead

Ex-Palantir project lead; UCL PhD. Carries the forward-deployment model the residency teaches.

500k+AI practitioners taught since 2019

15+in-house AI engineers, grown from our own line

21+AI applications shipped to production

10k+copies of Building LLMs for Production sold

08How it runs

Evidence before scale

01

Scoping call · 30–60 min

Team sizes, stacks, baseline, and target capabilities, discussed with the instructors. We'll tell you if Step 1 should come first.

02

Baseline & co-design

We assess and segment the developer population; a nominee per team keeps materials resonant with real work. Agreed with engineering leadership before anything is booked.

03

Pilot cohort

Typically 20 to 40 engineers, chosen with leadership. Exercises from your approved roadmap, never generic demos. Capstones build throughout.

04

Certify & scale decision

Certification against shipped capstones and a written debrief of what the cohort can now do. We scale to further waves only when the debrief earns it.

+Pilot first. One cohort, agreed success measures, and a scale-decision date, before any programme commitment.

+You own the artefacts. Capstones, standards, and reference builds stay in-house. No dependence on us for the next use case.

+Procurement-ready. NDAs standard, PO and invoicing supported, no production access required; your code stays yours.

+Portfolio roll-outs. For groups and PE portfolios: shared standards, per-company scoping, champions carried across companies.

The winners of the next three years convert their existing developer base fastest.

One call, one plan, one pilot. Scale only when the debrief earns it.

Book a scoping call Start with Step 1

09Logistics

Frequently asked questions

Do participants need to do Step 1 first?

It's the recommended order, not a hard rule. Step 1 builds the agent fluency this programme assumes, and its debrief tells you who to send here. If your developers already work seriously with coding agents, we confirm the baseline in scoping and start directly. Hard prerequisites: working Python and software engineering experience; a Python foundations route exists for mixed cohorts.

Where do capstones fit if delivery is part-time?

Cohorts run over weeks around delivery: self-paced study, live checkpoints, and capstone work that builds throughout. Certification happens when the capstone ships (a deployed, evaluated system), not when the calendar runs out.

Does "conversion" mean a guaranteed role change?

No. The programme develops applied AI engineering capability in existing software developers, certified against shipped capstone work. Scope, assessment, and outcomes are agreed per engagement.

What proof can we review before buying?

The curriculum is on this page, the underlying courses have public previews, our book is in print, and the production systems behind the curriculum are described above. References and the most relevant engagements are discussed in scoping.

10Next step

Book a scoping call

Tell us about your engineering team, current baseline, and the AI capabilities you want to build. We reply within one business day with a short scoping agenda and an indicative quote, and you'll talk with the instructors who deliver the programme.

1On the call: team sizes, stacks, baseline, target capabilities, and which formats fit which teams. 30–60 minutes, with instructors.

2You leave with: a recommended programme architecture, workload estimate, commercial range, and measurement plan.

3No pressure: if Step 1, or nothing, is the right answer, we say so.

Every enquiry is read by the Towards AI team. Prefer email? louis@towardsai.net