A personal AI that knows what you know.
Without you uploading any of it.
You already run your own stuff. NAS in the closet, Obsidian vault on disk, Tailscale between the devices, an allergy to one more subscription. The dream is an assistant that can answer over all of it — email, files, notes, photos, code — without surrendering it to do so. inseam is that layer, built for people who'd rather configure than rent.
Six things that work
because the index lives where you do.
The index doesn't copy your data — it indexes addresses and metadata. Your photos stay on the NAS. Your email stays in the mail store. Your notes stay in the vault. Your voice memos stay where they were recorded. That single decision is what makes the rest of this list possible.
The index is local. The summary tree is local. Embeddings are local. Your agent navigates the tree, finds the document, surfaces the summary plus enough context to answer the question without the body. If it needs the body, the request queues — when you land, fetch resolves through whichever Connection now reaches the source.
Ship inseam embedded in an iOS app. Sync a 100k-document personal corpus — email, files, voice memos, Notion, Linear. Total on-device footprint: a few hundred MB. The same corpus in any competing system's vector store is tens of gigabytes — they store the chunked content. They cannot ship on the phone that produced the data; they have to be cloud services by necessity, not by choice.
A 3B on-device model collapses when handed fifteen marginal chunks. inseam hands it a tight, hierarchically-scoped slice instead. Answer quality on the device matches what you'd get from a frontier cloud model plus a cloud RAG stack — at zero API cost, with zero data egress. You don't need a $20/mo API key for useful answers.
Apple Watch logs voice memos into a local micro-node. Your iPad runs the full app with Gmail and Drive Connections. Your Worker on Cloudflare is always-on and runs heavier embedding models. A query to any of them federates across all three — the watch's voice memos retrievable from the Worker, Drive files indexed by the iPad reachable from the watch, transparently.
You revoke Gmail's OAuth grant — maybe you're rotating credentials, maybe you're switching to Fastmail. Every Source row still exists, still searchable, still surfacing in retrieval. If another paired node holds Gmail credentials, retrieval transparently routes through it. When you re-auth weeks later, fetch resumes against the same addresses. No re-ingestion. No rebuilt index.
Homelab data lives in genuinely weird places. No vendor will ever ship connectors for any of it. The plugin contract is small enough that you (or an agent helping you) can write one in an afternoon, against framework-enforced fixtures and tests. Homelab forums become a connector ecosystem.