Go from scikit-learn scripts to shipping retrieval, agents, and evaluated language-model endpoints that survive real traffic.
You can train a model, tune it, and score well on a held-out set — and none of that tells you how to ship the feature your product actually needs. The models are already trained. The hard part is everything around them: choosing a checkpoint, turning documents into vectors, surfacing the right five paragraphs, and proving the answer wasn't invented. This course starts inside the transformer — you implement scaled dot-product attention yourself, not just read that it uses attention — then works outward: BERT versus GPT, HuggingFace pipelines and cheap fine-tuning, sentence embeddings and approximate nearest-neighbour search over a million vectors, retrieval-augmented generation (RAG) with hybrid search and re-ranking, tool-using agents that recover from failures, prompts versioned and A/B tested like code, evaluation harnesses that run in continuous integration, and production large language model (LLM) clients with retries, routing, and cost tracking. Every module is runnable Python plus a project you ship. The capstone is one real AI feature — semantic search, grounded question answering, or an agent — running in production with numbers attached.
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 applyFinished the tasks? Take the prompt to your AI and get tested on it. We copy the prompt and open the app — just paste it in.
Implement attention yourself so you can read any model spec, predict its memory cost, and explain why generation slows down.
A million checkpoints and most are wrong for your task — learn to choose one, fine-tune it on a small budget, and publish it back.
Keyword search misses the paraphrase your user typed; embeddings plus an approximate index find it across a million documents.
Retrieval-augmented generation (RAG) lives or dies on which chunks surface first, so you tune chunking, hybrid search, and re-ranking.
Give a model tools and it can act — until it loops forever or a tool fails. Build the loop, the stopping rules, and the recovery.
A prompt that shines in a notebook breaks at ten thousand calls — version it, defend it against injection, and price every token.
Without a harness, every prompt edit is a coin flip; with one, each release runs the same graded test set and fails loudly in CI.
Large language model (LLM) providers throttle, fail, and bill per token — build a client that retries, routes, and tracks spend.
Your real inputs are not all text — receipts, screenshots, and voice notes become structured data through one pipeline.
Complete 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.