Great research shouldn't just be read—it should be used. I see the open-source community as a living laboratory where real-world impact is the only metric that truly matters. Let's build in the open!
Open-source map
Lead
Lifelong personalized tutoring — an agent-native workspace where chat, quizzes, research and practice all run on one loop.
Django study assistant — notes, mind maps and practice questions.
Per-limb movement scoring from video, with local-LLM feedback.
What I build
DeepTutor treats tutoring as one agent loop rather than a collection of features. Chat, Quiz, Research, Visualize, Solve and Mastery Path are objectives on the same loop, so knowledge bases, notebooks, question banks and personas stay addressable from any of them instead of being siloed per tool. Retrieval is pluggable across LlamaIndex, PageIndex, GraphRAG, LightRAG and linked Obsidian vaults, and an external coding CLI can be consulted as a subagent mid-turn. The learner model is three explicit layers — L1 traces, L2 summaries, L3 synthesis — exposed through a memory graph, so every inference the system draws about a learner is inspectable and editable rather than latent in a prompt.
CatchMe argues that personal memory is a navigation problem, not a search problem. Six recorders capture GUI activity — window focus, keystrokes, mouse, screenshots, clipboard, notifications — and deterministic rules (window switches, idle timeouts, temporal clustering) fold them into a five-level tree: day → session → app → location → action. Structure is rule-based; content is LLM-summarised bottom-up as nodes close, incrementally rather than in an end-of-day batch. Retrieval carries no embedding index: a rule pass prunes by time, then an LLM traverses the tree, selecting multiple branches per level and reading raw evidence only where it needs to. That holds up on fragmented, cross-modal activity data where similarity search does not, in ~0.2 GB locally, behind a single skill file any CLI agent can call.