labs

P.I.P.E.R.

Private by design · open source · work in progress

The pipeline: every application as one path from source to outcome (sample data)

P.I.P.E.R. is a job-search co-pilot that runs entirely on your own computer. It reads your inbox, tracks every application, and is designed to draft your outreach and prepare your applications, but never to send or submit anything on its own.

An inbox holds bank statements, medical and tax mail and private conversations, not just job emails, and I didn’t want to hand that to an online service whose use of it I can’t see. So the model is local, the data stays on your machine, and the app connects to Gmail with read-only permission. The one other call is fetching a job posting, and only when you click “Fill from link”. No email text goes to an AI service, and there is no account to create. You can reach it from your phone through a private network such as Tailscale. It’s MIT licensed, so you can fork and host it yourself, and if there’s demand I could offer a paid cloud version with clear privacy and security assurances.

A job search is five jobs at once: knowing where you stand, writing to people, telling your story, preparing for conversations, and applying. Each lives in a different tool, or in your head. The idea is one assistant that carries the same context across all of them. The first module, with its analytics, is built and I use it daily; the rest are designed or planned, and I’m being careful to say which is which.

Pipeline
Reads your Gmail and tracks every application. Built, in daily use.
Analytics
How much you apply, to what, and what is working, plus a short summary written by the local model from calculated numbers only. Built.
Writing
Cover letters and messages, with every claim traced to its source. Fully specified.
Résumés
A tailored version per role. Planned.
Interviews
Likely questions and your best-matching stories. Planned.
Applying
Prepares an application for you to review and send. Planned.

A 7B model that fits an 8 GB graphics card is weaker than a hosted one, so when it isn’t sure about an email it asks you instead of guessing. The case study covers why I chose local and what it costs.

Role
Product designer and engineer, solo, end to end
Constraints
Fully local, one 8 GB consumer graphics card, read-only access to a personal inbox
Stack
Python, FastAPI, SQLite, React, TypeScript, Tailwind, ECharts, Ollama running Qwen 2.5 7B, Gmail API
Status
Inbox pipeline and analytics built, pipeline in daily use; open source on GitHub

Read more