Welcome to Braedyn Thompson's portfolio!
See the roles I'm targeting →

Now playing
Berkeley Lab SeismicSoCal BearLM
PROJECT 06 · DESKTOP APP · ENTITY RESOLUTION · LOCAL-FIRST

JobbyMail

A Windows app that reads your mailbox and keeps your job applications up to date by itself. It classifies each email, works out which application it belongs to with q-gram entity resolution, and derives every status from an append-only event log. The mail, the database and the models stay on your computer.

Read-only mail accessPython · SQLiteTauri · ReactRuns on your PC
matching 6,000 emails to 5,000 companies99.99%candidate pairs skipped100%follow-ups matched

Overview01

4statusesOffer, Interviewing, Applied, Rejected
99.99%of comparisons skippedby q-gram blocking, at 5,000 companies
0serverseverything runs on the user's PC
12 MBinstallerWindows, no admin rights

Applying to jobs produces a stream of email: a confirmation, then silence, an interview invite, a rejection, sometimes an offer. JobbyMail reads that stream and keeps the table a job seeker would otherwise maintain by hand.

A Python pipeline syncs the mailbox read-only over IMAP, classifies each message, extracts the company and role, and resolves the email to an application. Every change is an event in an append-only log in SQLite, and each application's status is worked out from its events, so running the pipeline twice changes nothing and a wrong guess can be corrected without losing history. A Tauri window with React pages shows the result.

  • 01

    Built an incremental, read-only mail sync (IMAP with verified TLS; the app password lives only in the Windows credential store) that fetches only new messages, de-duplicates by Message-ID, and stores one odd message in raw form instead of stopping.

  • 02

    Resolved each email to an application with entity resolution: names cleaned, candidates found by q-gram blocking, then matched on q-gram overlap across the company name, the sender's name and the sender's address. Blocking skipped 99.99% of 17.5 million candidate pairs.

  • 03

    Derived every status from an append-only event log: a status only moves forward on email evidence, a rejection applies from any state, and a status set by hand wins until a newer email arrives.

  • 04

    Rendered untrusted HTML email safely: an allow-list sanitiser, a sandboxed frame that cannot run scripts, and a content policy inside the frame, with remote images blocked until the reader allows them.

  • 05

    Wrote the tests during the build and ran them all at the end: 417 tests across four suites (Python, the React UI, the Rust shell and the browser) now pass; the few that failed on the first full run were faults in the tests, not the app.

Reported honestly: every accuracy and timing figure on this page was measured on invented emails written for the build. Real-mail accuracy is unmeasured until a labelled set of real messages exists. The app is Windows only, the installer is unsigned, and Outlook.com accounts are not supported. The code is in a private repository.

From an email to a row02

01SyncIMAP, read-only; only messages newer than the last one seen
02Classifyconfirmation, interview, rejection, offer, other job mail, or not job mail
03Extractcompany, role and date, from the platform's template, the sender or the text
04Resolvewhich application is this about? q-gram blocking, then matching
05Recordone event appended; the status is derived from the events

Each stage writes its own table, so any stage can be re-run without reading the mailbox again. The mailbox is opened read-only, so the app cannot mark, move or delete anything, and attachments are never stored.

A rule-based classifier that avoids false rejections

The first classifier is a set of rules. Marking a live application rejected is the costliest error, so a status phrase is ignored when the sentence is conditional ("if you are not moving forward…", "until the position has been filled"), a status needs a job word nearby, and a rejection that sits beside "we received your application" becomes a question for the user instead of a status change.

One odd message never stops a runA message with an impossible date or thousands of nested parts is stored with its raw headers and skipped past, so the sync position keeps moving.
Hiring platformsMail sent through an applicant-tracking system is recognised by the platform's name in the sender's domain, and the employer is read from the text.
Job boardsA confirmation sent by a job board names the employer in its subject or body; the board itself is never taken as the company.
A file of real labels stays out of the repoThe labelling command refuses to write inside the repository, so real addresses cannot be committed by accident.

Entity resolution03

"Acme Recruiting", "Acme Corp" and a hiring platform's address can all be the same employer. A later email has to find the right application, and the right role when there are two.

01Cleanlowercase; drop legal endings (Inc, LLC) and hiring words (Careers, Talent Team)
02Blocksplit into three-letter q-grams; keep only companies sharing enough of them
03Matchscore the survivors by q-gram overlap on three signals
04Roletitle similarity, then job number, then email thread, then the only open application
ScoreOutcome
0.70 or morematched to that application
0.40 to 0.70a question: “Is this email referring to the Data Analyst job at Acme?”
below 0.40a different company

The thresholds came from 40 hand-made name pairs: the same company scored 0.41 to 1.00 and different companies 0.36 to 0.60. The ranges overlap in the middle, and that band is what goes to the user as a question.

Short namesIBM or HP have too few q-grams to score, so a cleaned name under five letters needs an exact match or a matching sender domain.
Two roles at one companyAn application is a company and a role. A rejection that names no role while two applications are open is never applied to both; it becomes a question.
The name outranks the addressMany employers send through the same hiring system, so a shared sender address alone never merges two companies.
Sent mail countsAn application the user emailed directly has no confirmation; the recipient's address names the company.

At 5,000 companies and 6,000 emails, blocking compared 2,229 of 17.5 million possible pairs and the run took about 12 seconds, with all 1,000 follow-up emails matched to the right application.

A rule that changed during the build: at first every email in the middle band became a question. A scale test showed that two different companies sharing a common word ("… Robotics") land in that band, which would have turned every new application at a similarly named company into a question. A first confirmation from a merely similar name now starts a new application; interviews, rejections and offers still ask.

Events and status04

JobbyMail jobs table: company, role, applied date, a status select and an Archive button on each row
Screenshot The jobs table on invented data: offers first, then interviewing, applied and rejected. Two rows show an email that disagrees with a status set by hand, as a suggestion to accept or dismiss.

Nothing in the event log is ever updated or deleted; the database refuses both. Each event from an email carries a key built from the message and the application, so processing the same mailbox twice adds nothing.

Forward onlyEmail evidence moves a status from Applied to Interviewing to Offer. A later confirmation never demotes an interview.
ReopeningAfter a rejection, a later interview or offer email reopens the application; a later confirmation does not.
A person outranks the pipelineA status set by hand stays until a newer email arrives, and then the email shows as a suggestion to accept or dismiss.
Archive with a fuseArchiving starts a four-second fuse with Undo; the archive event is written only when the fuse ends.
JobbyMail review page with questions about which application an email refers to
Screenshot The review list: emails the pipeline was unsure about, each with a direct question and one-click answers.

Showing untrusted email05

Clicking a job opens the email behind it inside the app. HTML email is untrusted input, so it passes three layers before it is shown.

01Sanitisean allow-list built on the browser's own parser: scripts, forms, frames and event handlers removed
02Sandboxa frame with no script permission and no access to the app's own origin
03Policya content policy inside the frame that denies everything by default
JobbyMail email viewer showing a message with remote images blocked and a Show remote images button
Screenshot A deliberately hostile sample with a tracking image. The image is blocked and the reader is told why.

Remote images are blocked until the reader presses "Show remote images" for that one email, because loading an image tells the sender the mail was opened. In review, 37 hostile emails written for the purpose (script tags, event-handler attributes, javascript: links, forms, meta refresh, CSS url() tricks, SVG and MathML shapes) were each opened both ways: no script ran and no request left before the button was pressed.

In the desktop window, a link in an email opens in the system browser, and only for web and email addresses; the allow-list is in the window's Rust code, and the page itself has no permission to open anything.

Insights and a gated model06

JobbyMail Insights page: reply rate, a funnel chart, days to first reply and a silence curve
Screenshot The Insights page on invented data. Every figure states how many applications it rests on.

The Insights page answers what a job seeker asks of their own search: how many applications get any reply, how long a reply takes, and when silence starts to mean no. The silence estimate is a Kaplan-Meier style curve that treats applications still waiting as not answered yet, instead of counting them as never answered. Any rate or estimate needs at least 10 applications behind it; below that the page says "not enough data yet" and shows only counts.

A trained classifier behind a gate fixed in advance

A TF-IDF and logistic-regression classifier, written in the standard library so the installer stays small, can replace the rules, but only by passing a gate written into the design before any real result existed:

Enough labelsAt least 300 labelled real emails, with at least 20 rejections.
Better by a marginMacro-F1 at least 0.03 above the rules', on emails the model did not train on.
No more wrong rejectionsRejection precision not lower than the rules'.
No leakage between foldsCross-validation folds are grouped by sender domain, so one company's repeated template cannot flatter the model.

No model ships with the app. The mechanism was proven on generated emails, with sets built to pass and to fail the gate; a failing run removes any older model and the rules take over again.

The model file is JSON written and validated by the app's own code, never a pickle, so a corrupt or planted file cannot run anything.

The desktop app07

JobbyMail checklist page: a grid of job boards with check boxes and a weekly checklist below
Screenshot The checklist: pick the job boards you use, then tick each one off. It clears weekly or daily.

The window is Tauri, which uses the system's own web view, so the installer is 12 MB. The pipeline is frozen into a self-contained folder and started by the window from its own resource folder; it serves a local connection on 127.0.0.1 that requires a random token, and it stops when the window closes.

OnboardingA five-step stepper collects a name, a job field and target roles, connects the mailbox with a test button, and asks how far back to read.
One place for the passwordExactly one route accepts the app password and hands it to the Windows credential store. It never reaches the database, a response or a log.
Delete everythingA single action empties the database, the trained model and the stored credential, after the user types a word to confirm.
JobbyThe mascot is one of eight original vector sprites, each a glossy white space robot that waves; the chosen one (Orb) is a one-line change.
A grid of eight white robot mascot designs on black backgrounds
Screenshot The eight Jobby designs, each at mascot, profile-picture and icon size; Orb is in use.

How it was built08

The app was built in 15 phases by a loop of four roles: an architect that wrote each phase's plan and its checks, a builder, a checker that ran every check and never edited code, and the architect again as reviewer.

Tests written early, run onceEach phase wrote its tests without running them; fast checks (it compiles, it lints, it starts, a command prints the expected numbers) guarded each phase, and the whole suite ran in the last one.
A regression stops the loopAny check that had passed and then failed halted the build. It fired once, when a timing check ran on a machine short of memory, and the run was repeated on a quiet machine.
Reviews that rejectMost phases were sent back once: a classifier that read “if you are not moving forward” as a rejection, a typed date that landed a day early west of UTC, a password that could be left behind in the credential store.
Exact-match checksVendored interface components are pinned byte for byte to their published source, and the preview data is compared field by field with the real local API.

By the end the build carried 110 checks. The installer is produced by one command: the pipeline is frozen, the window is compiled in release mode and both are bundled into a per-user Windows installer.

Still open: accuracy on real mail, Microsoft sign-in for Outlook.com, real job-board logos in the checklist grid (lettered tiles stand in), and a signed installer.