Offline-first AI, built from the papers up
Airplane Mode Labs builds the tools that AI usually needs the cloud for, so that they run offline, in the browser, on small hardware. The first product is an offline document intelligence kit for Indian financial, legal and healthcare documents that works on an 8 GB laptop with the Wi-Fi off.
What is here
- Papers: the reading list, with each paper's claim, our status and notes.
- Benchmarks: measurements on modest hardware, or an honest "not measured".
- Write-ups: what a paper claimed, what held up, what did not.
Demos
- pagetriage: Sprint 5 (18 Nov 2026): Working: the same page analysis as before, now with the PDF structure call that the dockit demo uses documented (analyse_pdf_structure: which pages of a PDF hold a readable text layer), with its WebAssembly size and its timings measured in Chrome on a 48 GB Apple M5 Pro (not measured on 8 GB). Not yet: OCR, 8 GB figures (not measured on 8 GB), a signed release, a tag, a published package, Linux and Windows wheels (Sprint 6).
- runtime: Sprint 5 (18 Nov 2026): Working: runtime beta 0.3.0-beta.0 on spec release 0.5.0: capability profile, signed-manifest verification (now with an expected-id check that refuses a validly signed manifest of another product), download, hash check and quota eviction of synthetic assets, resume, and a Tamper lab of fourteen cases that shows the exact refusal reason, with 0 requests after a manifest refusal. The test-model page adds embeddings through the worker host, in-worker hashing of every model file, a token budget and Stop. This page's policy is now connect-src 'self': it can only reach its own origin. Not yet: WebLLM (cut), Windows and 8 GB measurements (not measured on 8 GB), key revocation, a real signing key.
- glyphsift: Sprint 5 (18 Nov 2026): Working: everything from Sprint 4, plus a Gate tab that gives a pasted chunk of Devanagari text one of three answers, pass, hold or not scored (text that is not mostly Devanagari is passed through and not scored), with the score, the Devanagari share and a threshold you can change (975000 ppm by default), as the same canonical JSON the command line and the Python wheel print, computed in WebAssembly in the tab; text with an unpaired surrogate is refused in plain text; 0 requests after the page loads. Not yet: scripts other than Devanagari (Tamil is planned for v0.2 as a design note only, Bengali later), a published package or wheel, 8 GB figures (not measured on 8 GB).
- localdata: Sprint 5 (18 Nov 2026): Working: a documents API (putDocument, queryChunks returning chunk text and source page, deleteDocument; all-or-nothing, with clear(ns) removing a document namespace physically), prototype-safe keys and values, size-checked OPFS reads, a sweep of unreferenced records on open, all on OPFS, IndexedDB and memory in Chrome on an Apple M5 Pro with 48 GB (not measured on 8 GB). Not yet: HNSW, quantisation, compaction, encryption at rest, an 8 GB measurement.
- dockit: Sprint 5 (18 Nov 2026): Working: ask a question about a PDF and read an answer from a small local model (Qwen3-0.6B, WebAssembly) with the source pages shown; page triage (pagetriage), text-layer extraction (pdf.js in its own worker), a glyph gate for Devanagari text (glyphsift) with the reason shown when a passage is held back, chunks and vectors in localdata, and a capability card; scanned, oversize, corrupt and encrypted files and a missing or tampered model are refused in plain words; after the one-time model download no request leaves the page (proved on a local build in installed Chrome on an Apple M5 Pro 48 GB, not measured on 8 GB; the live page changes only after the owner pushes). Not yet: OCR, a Windows or 8 GB figure, the pilot model (Qwen3-1.7B), a real signing key, WebLLM (cut), multi-document chat.
- site: Sprint 5 (18 Nov 2026): Working: the benchmarks page now carries the dockit end-to-end rows (time to first token, total time, peak memory as the Chrome process tree, and the JS heap and WebAssembly memory under their own names, with the fixed question-set result as N of M correct with the correct source page), built only from committed data, with the done-when line, a limitations list and an 8 GB column that reads not measured on 8 GB; every check that read a sibling repository now reads a committed fixture, so CI runs from a lone checkout. No scripts. Not yet: 8 GB figures (not measured on 8 GB), a Windows measurement, a published Sprint 5 write-up (a skeleton only).
Statuses written on 18 Nov 2026 from local builds, not measured at the live URLs. See the demos page.
Repositories
| Repo | Job | Stack | Licence | Status |
|---|---|---|---|---|
| pagetriage | What is on a scanned or PDF page, and where it should go next | Rust core, Python wheel, WebAssembly | MIT OR Apache-2.0 | Working; installable from a locally built wheel and npm tarball (macOS arm64); not published, no tag yet |
| glyphsift | Is extracted Indic text clean enough to trust or train on; Hindi first | Rust core, WebAssembly, Python wheel, CLI | Apache-2.0 | Release candidate 0.1.1 (not tagged, not published): validator, gate call, CLI, WebAssembly demo with a Gate tab |
| localdata | set/get/delete/query over OPFS with an IndexedDB fallback | TypeScript SDK; flat vector index scan in Rust compiled to WebAssembly | Apache-2.0 | 0.3.0-alpha.0 (not tagged, not published): set/get/delete/list, query() over a flat vector index and a documents API (putDocument, queryChunks) persisted in OPFS (not measured on 8 GB) |
| runtime | Capability profile, asset manager, backend adapter, observability events; the open specs | TypeScript | Apache-2.0 | Beta 0.3.0-beta.0 on spec release 0.5.0 (not tagged, not published, not tag-ready): capability profile, asset manager, backends, worker engine with embeddings, DevTools panel; WebLLM: not yet (cut) |
| dockit | Reference app: offline PDF question answering as an installable web app | TypeScript, React, PWA | Apache-2.0 | 0.2.0-dev (private, not published): answers a question about a PDF offline with a small local model; OCR: not yet; not measured on 8 GB |
| site | This site: paper tracker, benchmarks, write-ups | Plain HTML, zero dependencies | Apache-2.0 code, CC BY 4.0 prose | Live |
| control-plane | Signed registry, offline update bundles, policy, audit, fleet view (private) | Rust | FSL-1.1-Apache-2.0 | Private; manifest and signing crates only |
Package names use the placeholder scope @airplanemode and crate names <repo>
until the handles are reserved; nothing is published yet.
How we work
- Papers before tools
- We read the key papers for a component before we look at how others built it. Existing tools are something to measure against, not something to copy.
- Small repositories with one job each
- Each tool gets its own repository, a clear purpose and a README with a benchmark we are willing to defend.
- The browser and small hardware are the default
- WebAssembly, WebGPU, OPFS and IndexedDB first. Anything that needs a server has to justify itself.
- Measure, and publish what did not work
- A shared harness reports memory, latency, size and quality on modest machines. Negative results are published alongside positive ones.
- Ship something small every week
- A small public artefact every Saturday beats a large private one at the end of the quarter.
- Write it down
- Every reimplemented paper gets a short note: the claim, what held up, what did not.
Everything on this site runs without JavaScript and loads nothing from a third party.