917: 8 Steps to Becoming an AI Engineer, with Kirill Eremenko

Episode Summary

Executive Summary: Kirill Aramenko outlines an eight-week AI engineering boot camp that teaches participants to move from AI proof-of-concept to production-grade systems. The first half focuses on model selection, prompt/behavior design, RAG, and agents; the second half emphasizes deployment on AWS, caching, memory, security, observability, and capstone productization. He frames AI engineering as business-first, emphasizing practical trade-offs over hype.

Main Topics: Boot camp structure and prerequisites: The program is designed for intermediate-to-advanced learners who already know Python, basic PyTorch/scikit-learn/pandas, LLM API usage, and AWS/cloud fundamentals. It runs in interactive weekly sessions and office hours, with small cohorts and feedback loops. Week 1: Mindset shift and model selection: Participants learn to avoid overbuilding with generative or agentic AI when simpler solutions suffice. They compare many LLMs and reasoning vs. chat-style models, focusing on defining business goals and success metrics first. Week 2: Behavior design and prompt engineering: The focus is on designing LLM behavior through system prompts, templates, and response parsing, using Gradio to build a simple flight assistant app. Kirill stresses that system prompts are reliably sent to the model. Week 3: Retrieval-Augmented Generation (RAG): Learners build RAG pipelines from scratch using chunking, embeddings, vector databases, and retrievers. The episode emphasizes hierarchical RAG, context-aware chunking, and the shift from fine-tuning to inference-time customization. Week 4: Agentic AI: Participants build agents that use tools, memory, and RAG, including a digital twin that can answer interview questions about a participant’s experience. The episode clarifies that agents orchestrate tools via text-based LLM interactions rather than directly acting on systems. Weeks 5-8: Production, security, and capstone deployment: The second half covers AWS deployment, Docker, Lambda vs. EC2 trade-offs, caching, memory, security testing, observability, and production RAG. The capstone turns the digital twin into a scalable, secure, CI/CD-driven product.

Key Arguments: AI engineering must start with the business problem; if goals and success metrics are unclear, do not build yet. Reasoning models and chat models serve different purposes: reasoning models reduce errors through slower, stepwise deliberation, while chat models generate faster stream-of-consciousness outputs. RAG remains essential even as context windows grow because it improves relevance, accuracy, and cost efficiency compared with stuffing everything into context. Agentic systems are not magical autonomous actors; they are LLMs orchestrating tools through text exchanges and code-managed tool routing. Production readiness is a core AI engineering skill: deployment, caching, scalability, security, and observability matter as much as model quality. Security should not rely only on prompts; sensitive actions need code-level constraints, logging, and verification of tool calls. AI projects should be treated as products that evolve over time, with memory and feedback loops to close knowledge gaps.

Data Points: Boot camp length: 8 weeks - The AI engineering boot camp is organized into eight weekly modules. Cohort size: 10 people per cohort - Kirill notes the boot camp has limited spots and small cohorts. Core weekly live session: 3 hours on Monday - The main learning/practice session is held on Mondays. Office hours: 2 hours on Thursday - Thursday sessions focus on discussion, use cases, and follow-up questions. Employer learning budgets: $5,000–$10,000 per employee per year - Kirill cites typical corporate training budgets as a way companies can fund the boot camp. Interview success timeline: Within 3 months - One participant aimed to land an AI engineering job within three months after the boot camp. First cohort status: Inaugural cohort ending in 2 days - Recorded near the end of the first cohort. Second cohort launch: 29th of September - Kirill mentions the next cohort start date at the end of the episode.

Pivotal Quotes: "If you can't define the business goal and success metric, don't build it yet." — Kirill Aramenko: Week 1 pro tip on starting AI projects with a business-first mindset. "The system prompt is always guaranteed to go to the LLM." — Kirill Aramenko: Week 2 discussion of behavior design and why system prompts matter. "Treat your agents like a product. It's not just a project, it's a product." — Kirill Aramenko: Week 8 capstone emphasis on continual improvement after deployment.

Implications: For learners and teams, the episode argues that effective AI engineering combines model literacy with deployment, security, and product thinking. Organizations should train staff to build useful, cost-aware systems that solve real problems, not just impressive demos.