01
AI Chatbot
A multi-turn conversational AI with streaming, a persistent system prompt, configurable context windows, and real-time token/cost tracking — you build the conversation loop yourself, not a wrapper around a chat UI.
Available now
02
Token & Embedding Explorer
An interactive tool that visualizes how text gets tokenized before any LLM sees it, and how embeddings represent meaning as vectors you can compare by similarity.
03
Semantic Search
A semantic search engine over a corpus: embed documents, store them in a vector database, and retrieve the most relevant results by meaning — with no keyword matching.
04
PDF Research Assistant
A RAG system that ingests PDFs, chunks and indexes them, retrieves relevant sections, and generates grounded answers with citations — plus a faithfulness check that detects hallucination.
05
Personal Memory System
A persistent memory system that gives an AI assistant memory across conversations — scoring memories on relevance, recency (exponential decay), and importance, across episodic (what happened), semantic (what is true), and procedural (how to do things) kinds.
06
AI Coding Copilot
A context-aware coding copilot that reads your codebase through tools — file read, directory listing, code search — and answers grounded in the real files, not a guess. The learning target is tool use and the agentic loop (ReAct): the model requests a tool, you execute it, feed the result back, and repeat until it answers.
07
AI Evaluation Framework
A reusable evaluation harness for measuring AI system quality: LLM-as-a-judge scoring, a pass-rate decision, and regression testing — re-score a frozen dataset after a prompt/model change and catch what got worse. The learning target is the judge (prompt + robust parse) and turning scores into decisions, with its biases (position, verbosity, self-enhancement) reproduced as failure analysis.
08
AI Agent
An autonomous, multi-step agent built by making Project 06's tool loop reliable: loop/stuck detection, a steps-and-tokens budget, tool-error recovery, and run evaluation (completion, efficiency, cost) — plus the judgment of when not to use an agent at all. The tool layer is provided; the learning target is the reliability layer that keeps an autonomous loop from going wrong.
09
Personal Learning OS
The capstone: a personal AI operating system that orchestrates memory, retrieval, the agent, and evaluation behind one routing front door for personal knowledge management. The direct prototype for StarcallOS patterns.
Electives
Optional, off-spine labs that go deeper on a focused topic. Not part of the numbered 1–9 sequence and they gate nothing — take them after the prerequisites listed on each card.
E1
MCP Interface Layer
Stop hand-wiring tools into one app loop. Wrap Project 05's memory system as a Model Context Protocol server — tools, resources, and a prompt — then consume that one server from two different hosts. The learning target is the provider/consumer split (M×N → M+N): turning a subsystem into a reusable protocol surface, with a real trust boundary. The backend and transport wiring are provided.