A living-room assistant. It listens through a Xiaomi camera mic, wakes on "Zosiu", and answers by voice and with pages it draws on the TV.
Python
MCP
OpenRouter
go2rtc
Android TV / Cast
Ear → agent → TV loop: ffmpeg + Voxtral wake word, agent over MCP tools, one Cast session for picture and sound
Tools: look and move the camera, show and edit pages, lists, timers, reminders, calendar, TV remote
Memory modelled on Hermes Agent: short notes in every chat, FTS5 trigram history, background review after silence
Skills: cooking, morning overview, charts and diagrams, fact checking, Netflix / YouTube deep links, films from a Filmweb DB
Write up the architecture: the mic is the loop input, not a tool; the model never gets "say"
Measure latency end to end and pick the cheapest model that still feels instant
Home & robots2026
TelloEye
● activepush31 commits
Teaching a toy drone to fly from a single camera: monocular depth → point cloud → navigation, including "fly to the TV".
Python
djitellopy
MiDaS
YOLO
Open3D
2.5D simulator with the djitellopy API, photo textures, per-pixel ground-truth depth
Swappable depth estimators (GT, heuristic, MiDaS, Depth Anything) and point-cloud fusion to .ply
Autonomous VFH-lite flight, FPV game with a REAL ↔ DERIVED view, object-nav autopilot
Real Tello adapter and camera / depth calibration (no flights yet)
Found out why: classic SfM degenerates on a 360° spin in place (zero baseline)
First real flight: SimTello → Tello with live view, height-based depth scaling, emergency stop
Benchmark Depth Anything V2 against MiDaS
Online voxel map from the estimated cloud instead of ground-truth occupancy
pytest suite instead of one-liners
Home & robots2026
Roślinki
‖ pausedkeep
iPhone app: snap a plant, get a guess (flower, tree…) from my own Worker API.
Expo
Sign in with Apple
Cloudflare Workers
Search & memory2026
DigUp
● activepush
Contributing to an open-source Mac app that finds files by what is in them (photos, video moments, PDF clauses) with an on-device multimodal embedding model.
Fork of ARahim3/DigUp.
Swift
EmbeddingGemma 2
Accelerate
MCP
Photo search: filter by when a photo was taken, browse by time or person with nothing typed
Better ranking: return nothing when nothing is close, fold far-behind matches
"digup mcp": serve the index to AI agents
Vector math through Accelerate on macOS, plain loops elsewhere
Split into a portable DigUpCore, the macOS app, a CLI and the MCP server
Upstream the photo work
Search & memory2026
xhose
‖ pausedmerge → DigUp / one personal index
Pull my X bookmarks out (scraper or extension) and search them by meaning in a local web page.
TypeScript
Playwright
Chrome extension
SQLite
in-browser embeddings
Search & memory2026
Highlight → Markdown
‖ pausedmerge → one personal index
Highlight text in Chrome, tag it, save it as Markdown with URL and date; qmd makes the pile searchable.
Python
uv
Chrome extension
qmd
Search & memory2025
Google Autocomplete, reversed
✓ shippedpush
What Google's autocomplete API tells you without meaning to: subtype codes reveal brands, Knowledge Graph entities, local businesses, maps intent.
TypeScript
Python
Search & memory2025
Primer
◌ ideakeep
Ask "how do you build a search engine?" and get a Wikipedia-style intro page: who works on it, their blogs, companies, papers.
TypeScript
LLM agents
JSON
Games2022 · 2026
Letters from Whitechapel, solved
✓ shippedkeep
The real 195-circle board as a graph: optimal hideouts, where to put the wretched, a worst-case road home. Playable on this blog.