Course Catalog

Course Description

Beyond simple text-to-image. ControlNet: conditioning on edges, depth maps, poses, segmentation maps. LoRA for image generation: concept fine-tuning, style transfer, character consistency. IP-Adapter and image-conditioned generation. Inpainting, outpainting, and img2img workflows. ComfyUI node-based pipelines. Quality evaluation: FID, CLIP score, human preference models. Students build complete production generation pipelines.

Course Description

Foundations of information retrieval from first principles. The Boolean retrieval model and its limitations. TF-IDF: term frequency, inverse document frequency, and their variants. BM25: the probabilistic model that powers most search engines today. Inverted indexes: construction, compression, intersection algorithms. Evaluation metrics: precision, recall, NDCG, MRR. Static site search with Pagefind. Elasticsearch and Solr: practical configuration for content search. Students build a complete search system for a document corpus.

Course Description

Understanding and improving user queries. Classic query expansion: pseudo-relevance feedback, thesaurus expansion, WordNet. Neural query expansion: using LLMs to generate synonyms, related terms, and alternative phrasings. The Scolta approach: LLM-based query expansion that improves BM25 recall without vector infrastructure. Query classification: navigational, informational, transactional. Spelling correction and fuzzy matching. Query suggestion and autocomplete. Students implement and evaluate a query expansion system, comparing classic and neural approaches on the same query set.

Course Description

Using LLMs to synthesize search results and answer user questions. AI Overviews (formerly Google SGE): architecture and user experience. RAG for search: retrieve-then-generate, citations, grounding. Scolta's AI overview feature: implementation deep-dive, prompt design, caching strategy. Hallucination in search contexts: detection and mitigation. Attribution and citation generation. Evaluating AI-generated summaries: faithfulness, relevance, coherence. Streaming responses for low-perceived latency. Students implement an AI overview system for a content corpus.

Course Description

The capstone course for the Certificate in AI-Powered Search. Students design and deploy a complete AI-enhanced search system for a real or realistic content corpus. Requirements: query expansion (LLM-based), semantic reranking (cross-encoder), and AI overview generation. Students choose their stack: Scolta on Drupal/WordPress, Elasticsearch + LLM layer, or a custom build. Final deliverable: a working system with evaluation report comparing baseline BM25 search against the AI-enhanced system on a curated query set.

Course Description

This foundational course covers the transformer architecture in depth, from mathematical first principles through efficient implementation. Students gain the understanding needed to read and extend modern LLM codebases, not just use them. Weekly labs implement key components from scratch in PyTorch.

Learning Objectives

• Derive scaled dot-product attention from information retrieval first principles
• Implement multi-head attention, residual connections, and layer normalization
• Analyze the computational complexity of attention and its practical implications
• Compare architectural variants: encoder-only (BERT), decoder-only (GPT), encoder-decoder (T5)
• Understand Flash Attention, grouped query attention, and memory-efficient implementations

Course Description

Tokenization is the first and often most overlooked decision in building a language model. This course examines how tokenization shapes what an LLM can learn, what arithmetic it can perform, and how well it handles multilingual text. Students implement BPE from scratch, analyze tokenizer behavior across languages, and understand the direct connection between vocabulary choice and model capabilities.

Key Topics

• BPE algorithm: merge rules, vocabulary construction, encoding and decoding
• WordPiece and SentencePiece: differences and use cases
• Tokenization in Llama, GPT-4, and multilingual models
• Tokenizer fertility: how tokenizer choice affects different languages
• The Feste tokenizer (Rust implementation): a walkthrough of production-quality BPE

Course Description

Training a large language model from scratch requires coordination across data engineering, distributed systems, and optimization. This course covers the full pretraining workflow: from assembling a high-quality dataset through designing the training run and diagnosing instabilities. Students analyze real pretraining logs from open-source models and work through case studies of successful and failed runs.

Course Description

Most practitioners never train a model from scratch — but almost everyone fine-tunes. This course covers the spectrum of adaptation methods from full fine-tuning through highly parameter-efficient approaches, with emphasis on practical tradeoffs: compute cost, memory, downstream performance, and catastrophic forgetting. Students fine-tune models ranging from 7B to 70B parameters using LoRA and QLoRA.

Course Description

Alignment transforms a capable language model into one that is helpful, harmless, and honest. This course examines the technical mechanisms behind modern alignment: RLHF, its practical challenges, and newer approaches like DPO that achieve similar results more efficiently. Students implement a full RLHF pipeline on a small model and analyze alignment failures in case studies.

Course Description

Getting an LLM into production efficiently requires understanding the full inference stack: from quantization and caching to serving frameworks and cost management. This course equips students to make informed engineering decisions about LLM deployment — achieving target latency and cost goals while maintaining acceptable quality degradation.

Course Description

Agentic AI systems are transforming software engineering. This course covers the full engineering stack for building autonomous agents that use tools, maintain state, plan multi-step tasks, and collaborate with other agents. Students build three complete agent applications over the semester, culminating in a multi-agent system capable of completing complex, open-ended tasks.