AriOS

Develop · Build

How it came together.

1. 01

   Telegram as the OS, not iOS Shortcuts

   I started with iOS Shortcuts for photo capture and a Chrome extension for URLs. Two surfaces, two auth flows, two failure modes. Telegram collapses both: one bot reachable from any phone, laptop, web client, or iPad, with a native share sheet from every iOS app. Long-poll the Bot API, allowlist my user ID, and the bot becomes a universal inbox for photos, voice notes, documents, and links. Everything routes through one /ingest endpoint on the server. New surfaces ship as thin clients.

   Backend Engineering
2. 02

   Heuristic classify is the default, not the fallback

   Ollama on a Dell CPU takes 30 seconds to classify a 1500-char prompt into a 12-option enum. That's unacceptable on the happy path. So the classifier inverts: a phrase-aware heuristic strips clickbait separators, clause-boundary truncates titles at 4-7 words, and matches a 12-domain regex against the body. Right answer about 80% of the time, in under 50 ms. The LLM only fires when the daemon is responsive AND the heuristic is unsure, which means most ingests are fast and the rare ones get the smart path. Failure modes inverted: nothing waits on a slow model.

   AI & Machine Learning
3. 03

   Vision moved to Cloudflare after every local tuning failed

   I spent a session tuning llava-7b on the Dell, sharp pre-resize to 672 px, num_predict 220, num_ctx 2048, 30-minute keep-alive, P-core-only threads, trimmed prompt. The baseline was 90 s and the timeout was 90 s. After every lever was pulled it was still 162 s per image because vision-token tokenisation dominates prompt-eval on CPU, not text generation. Anthropic was off the table on principle. Cloudflare Workers AI runs llava-1.5-7b on GPU with a free tier covering 10k requests a day; warm round-trip is 10-15 s. The vault stays local, only the photo bytes leave the device, and only for the perception step. The provider layer is one env var to flip, so I can move back to local the day a GPU mini-PC lands.

   AI & Machine Learning
4. 04

   Voice transcribes locally, photos route to cloud

   Voice notes are 1-2 minute Whisper-base ONNX inferences via @xenova/transformers, slow enough on CPU that I had to add a 'transcribing…' ack so the user knows the bot heard them, fast enough that the vault entry is written before they switch apps. Photos go to Cloudflare because vision is the harder local problem. The split is honest: keep what works locally local, send what doesn't, never pretend either is the full answer.

   AI & Machine Learning
5. 05

   The bot writes its own resilience

   Telegram's long-poll dropped silently after a bad socket and the bot got stuck in a 5-second-retry loop for hours. Fix: AbortController on every poll with a 35 s ceiling, exponential backoff with a 60 s floor, and a debounced undici dispatcher reset that fires on any consecutive failure across both the poll and the localhost forward path. The same dispatcher handles both, so a borked pool can starve both surfaces at once. Now a network blip clears in under a minute without intervention.

   Backend Engineering
6. 06

   Vault RAG over the same retrieval index

   There's no separate database. The retrieval index is an append-only JSONL file next to the vault: BM25 tokens + 384-dim MiniLM text embedding + 512-dim CLIP image embedding per entry, fused via Reciprocal Rank Fusion. Ask the bot 'what did I save about the SEO trick' and it BM25-plus-cosine-searches the vault, hands the top 4 passages to Ollama as context, and answers with [1] [2] citations. When Ollama is slow it degrades to retrieval-only mode, the user still gets the relevant sources, just without prose. The index can be rebuilt from the markdown at any time; the vault is still the source of truth.

   Data Pipelines