Threadline

// THE IDEA

One tree that both searches and explains.

Most AI assistants for mail run two engines: a vector database to search, and a language model to answer. You can't see inside either one.

Threadline files every email into a tree instead: category → topic → thread. The same tree is what it searches when you ask a question, and what you can browse yourself when you'd rather not ask at all. No embedding model, no vector database, and nothing hidden.

Mailbox 142 THREADS

Lectures & Profs

Placements

Northwind Analytics

Summer internship drive 3 EMAILS

Shortlist for the technical round

Bluepeak Systems

General College

Account Security

Ask "When is the Northwind interview?" and the answer comes from the highlighted thread, with the path to it lit up.

// HOW IT WORKS

From inbox to answer, in five steps.

Fetch

Gmail and Slack feed the same pipeline. Nothing downstream cares where a message came from.

Classify

An AI guess, checked by a deterministic keyword guardrail before anything is filed.

Thread

Replies join their conversation using real email headers, the way mail clients do. Re-sends collapse instead of piling up.

Store

Each thread lands in the tree with a running summary, urgency flags and its to-dos pulled out.

Ask

Questions are ranked against the threads, the best few are read in full, and the answer is grounded in that mail.

// THE GUARDRAIL

Every email gets two opinions.

A fast, cheap AI model makes a guess. Separately, a hand-built keyword scorer reads the subject and body. The scorer only overrules the AI when its own signal is clearly ahead, and every manual correction teaches it, so the same mistake doesn't happen twice.

A REAL EMAIL IT CAUGHT

"You shared some Google Account data with a third-party app"

AI GUESS"Cybersecurity"

WHY IT'S WRONGNot a category that exists

FILED ASThe right category, by the guardrail

// VECTORLESS RAG

Retrieval you can actually inspect.

TYPICAL RAG

An embedding model to host and keep from drifting
A vector index to build and keep refreshed
Similarity matches that are hard to explain

BM25 keyword ranking, no embedding model at all
A SQLite entity graph: category → topic → thread
Every match traces back to a keyword or an edge in the tree

// BUILT TO BE CHEAP

The cheapest AI call is the one you skip.

Skip when sure

If the keyword signal is already confident, the AI classification call never happens.

Skip for regulars

A sender with an established topic goes straight there, no AI needed.

Cache-friendly prompts

Mail is ingested in category order and prompts put stable content first, so repeated calls reuse the same prefix and providers' prompt caching kicks in.

Offline mode

One switch runs the whole pipeline with zero API calls: urgency, to-dos, filing and search all still work.

// EVERYWHERE YOU ARE

One backend, two apps.

Web console

React front end with a D3 mind map of the whole tree, analytics and an integrations page.

Android app

Built in Flutter, phone-first, with a full voice mode. Same account, same tree.

Calendar-aware

Google Calendar events are pulled into answers, so "am I free for that interview?" has a real answer.

Urgency and to-dos

Deadlines and "submit by Friday" lines are flagged by fast rules, at no AI cost.