Bridge from advanced ML to transformer-based AI systems. Embeddings, RAG, agents, eval, and production AI APIs — 100 challenges.
You know how to train a model. Now learn how to build with the models that already exist. This course starts with the transformer architecture (not just 'it uses attention' — you'll implement attention from scratch), then moves through BERT vs GPT, HuggingFace pipelines, sentence embeddings and semantic search, fine-tuning vs prompting decisions, RAG systems, AI agents with tool use, prompt engineering at scale, eval harnesses, production LLM APIs, and multimodal systems. Every module has runnable Python code and a real project to ship. The capstone is a full AI feature — search, or Q&A, or an agent — running in production.
Built by Lakshya Kumar
We grant free access case-by-case — students, career-switchers, builders on a tight budget. Sign in to send us a note.
Sign in to applyComplete all modules, then submit the required number of capstone projects. Each must earn a passing rating from an admin reviewer.
Pick one: (a) a semantic search API over a real corpus (≥1K documents) with hybrid retrieval and Recall@5 ≥ 0.75; (b) a RAG-based Q&A chatbot with a custom eval set of 30+ questions showing ≥70% answer faithfulness; or (c) a tool-using AI agent that completes a multi-step task (e.g., research + summarise + write draft email). Ship as a GitHub repo with: working Docker setup, README with architecture diagram, eval results, and a 5-minute video walkthrough.
Paste this into any AI chat. Fill in the bracketed parts with your context — you'll get back a straight answer on whether this belongs on your plate.
I'm considering a course called 'Applied AI: From ML to Modern Systems'. It bridges advanced supervised ML and modern transformer-based AI: transformer architecture from scratch, BERT vs GPT, HuggingFace, embeddings + vector search, RAG pipelines, AI agents, prompt engineering, eval harnesses, production LLM APIs, and multimodal. 100 Python challenges, capstone is a shipped AI feature. Context about me: 1. My ML background: [e.g. "I know sklearn but never touched a transformer", "I can fine-tune BERT but don't understand attention", "I build with OpenAI APIs but never trained anything"] 2. What I build day-to-day: [e.g. "backend APIs", "data pipelines", "Jupyter notebooks", "product features with LLM APIs"] 3. What I want to walk away able to build: [e.g. "a production RAG system", "an AI agent for my product", "a semantic search engine", "eval infrastructure for my AI team"] Answer: - Which 2 modules give me the highest leverage in the next 3 months? - What concrete artifact will I build that I can show in a portfolio or use at work? - Is this course right for me or should I do the Zero-to-Production ML course first? - What will I NOT be able to do after this — e.g. "train a custom LLM", "achieve state-of-the-art benchmark results", "replace a full ML research team"?
Build an evaluation harness for a RAG system: 50+ test queries with reference answers, automated retrieval-quality (recall@k) and generation-quality (faithfulness, relevance) metrics. Use it to compare two retrieval strategies (BM25 vs dense vs hybrid) on the same dataset.
Build a pipeline that extracts structured data (JSON conforming to a schema) from unstructured documents at scale. Include schema validation, retry-on-failure with progressive prompt fixes, and a 1000-document evaluation showing >95% structured-output success rate.
Build a function-calling agent that handles a domain task (booking, support triage, data lookup) using tool use. Include a tool registry, error-handling for tool failures, conversation state, and a 50-task eval showing it solves the task at >80% rate without human intervention.
Build a model router that sends each request to the cheapest model that meets a quality bar. Use evals to derive routing rules per task type. Demonstrate 60%+ cost reduction vs always-use-best while maintaining ≥95% of the best-model quality on a held-out test set.
Free interactive course. Pairs well with Modules 2–3.