Building AI Systems that hold up in production.
A five-week live cohort with Lan Chu for experienced engineers and architects responsible for AI systems that have to work beyond the demo, from retrieval and agents to platforms and evaluation.
AI systems fail in production for reasons the demo never showed.
A prototype that answers well in a demo meets harder conditions in production: documents that change, queries that get more complex, agents that loop, inference costs that grow, and users who trust answers they shouldn't.
The difficult part is rarely picking a model. It is deciding how data gets processed and retrieved, how much autonomy an agent gets, what the platform should centralize, and how the team will know the system is doing its job.
Across five weeks, work through those decisions with experienced engineers from other companies, then apply the thinking to AI systems you already build or operate.
Full syllabus
Becoming an AI-Native Engineering Team
Define what AI-native engineering means for your team today and how it changes product thinking, the SDLC, architectural trade-offs, engineering roles, and daily habits. Use frameworks from Patrick Debois, Ian Thomas, and Catherine Weeks to separate AI-augmented habits from AI-native ones and to keep risk management and resilience practices in place while adopting an agentic SDLC. Between sessions, map your current SDLC, propose where AI tools act as team members, name one part where AI deliberately stays out, and define a signal that shows the strategy is working.
Designing & Building RAG & Context Pipelines
Treat RAG as a data and retrieval problem as much as a model problem. Work through data processing, retrieval and reranking, knowledge graphs, and ephemeral versus long-term memory. Use Lan Chu's QCon London case study, Reliable Retrieval for Production AI Systems, alongside work from Adi Polak and Cassie Shum, to compare vector RAG and Graph RAG across contrasting use cases. Between sessions, design a retrieval architecture for a complex query at your organization, justify each layer of the pipeline, and explain what happens when the underlying data changes.
Designing & Building AI Agents
Design agentic systems that hold up outside a demo. Cover workflow patterns, orchestration, control, and memory, and critique an agent architecture for the three places agents most often fail: ambiguous instructions, tool errors, and runaway loops. Weigh fixed workflows against autonomy and single-agent against multi-agent designs, drawing on Itamar Friedman, Jake Mannix, and Adrian Cockcroft. Between sessions, design or prototype an agent for a real task, with a tool inventory, anticipated failure modes, guardrails, and an evaluation plan.
AI Platforms & Infrastructure
Work through the platform and infrastructure patterns that make AI systems production-ready: inference gateways, GPU cost controls, hosting choices, and what to centralize versus federate. Review a system architecture for inference waste using Meryem Arik's work, then decide which gateway policies your organization needs first, from model routing and rate limits to cost caps and PII handling. Between sessions, outline an internal AI platform strategy that separates batch from latency-sensitive workloads, drawing on Merrin Kurian and Joseph Stein.
AI Operational Excellence: Evals, Trust & Reliability
Examine the evaluation metrics and rollout readiness that make AI systems dependable in production. Most of the session is group capstone presentations and peer discussion.
Get the cohort details in your inbox.
We’ll send you the schedule, facilitator, full syllabus, weekly time commitment, price, registration link, and manager approval page in one email.
A live certification cohort, not another self-paced course.
Selected QCon and InfoQ technical material provides a starting point. The live sessions are where the group examines trade-offs, challenges assumptions, and sees how the same issue changes in other environments.
Between sessions, put the thinking into practice against systems, proposals, controls, assessments, or decisions you already own.
Lan stays with the cohort across the five weeks so later discussions can build on what participants tried between sessions.
How the five weeks work
Compare the same decision across other environments.
An approach that works in one environment can behave differently when the systems, constraints, team structure, or organization around it change.
The cohort gives you other reference points while the decision is still current. The value is not generic networking. It is seeing where an approach transfers, where it changes, and why.
Finish with an assessment in AI Engineering Program you can point to.
In Week 1, you choose a capstone. In the hands-on project, your group either designs a new AI-powered feature your organization would realistically ship (Track A), or takes an existing AI feature that works in a demo but isn’t production-ready and designs what it takes to close the gap (Track B). Both apply Weeks 1–4 across the full stack: SDLC boundaries, retrieval, agent design, platform stance, evaluation, and security. Alternatively, a group of 3–4 writes a 2,000–3,000-word technical article for publication on InfoQ. In Week 5, each group gives a 20-minute presentation followed by 30 minutes of peer discussion.
Complete the current attendance, participation, and capstone requirements to earn the InfoQ Certified AI Engineering Program credential.
Is this the right cohort for the work you own?
Designed for experienced practitioners, generally with at least five years of professional experience.
Start from QCon and InfoQ technical material.
Selected QCon talks and InfoQ technical material provide the starting point for the cohort. The live sessions are where participants examine those ideas against different systems, organizations, and constraints.
Open the shared Convince Your Manager page with this cohort preselected. The guided prompt uses the current current intake facts and asks about your role, current work, the problem you want to work through, what your manager cares about, and the most likely objection.
It then drafts a ready-to-send manager email, a version under 150 words for Slack or Teams, five short talking points, and one note on any missing detail that would materially strengthen the request.
Cohort questions
Building AI Systems that hold up in production.
with Lan Chu
Sign up for event updates
Hear about the latest InfoQ developments. Be the first to learn about our in-person and virtual events.