All projects

Case study 05 of 07 · LLM · Browser automation · Data quality

Personal AI Job Hunter

A local, single-user system that turns job hunting into a guided workflow: search, remove duplicates, rank, prepare with AI, fill the form in a real browser, and leave the last step to a person.

In daily personal use (local)Gemini for applications

In 20 seconds

What it is
A private tool that searches through official APIs and public feeds, ranks vacancies with an explainable scorer and uses Gemini to write grounded cover letters and answers. A Playwright queue then opens each employer form with details filled in.
My part
Product owner. I defined the product and its rules, made the architecture decisions (human in the loop, no sign-up platforms), and ran it against real job sites daily, reporting the defects that drove most of the accuracy work. Claude Code wrote the code.
Built with
Claude Code, Git and GitHub, pytest, VS Code
AI in the product
Google Gemini writes cover letters and application answers for the jobs I choose. Ranking is deterministic, not AI.
Status
In daily personal use (local). Specification on 8 Oct 2026 to a working system in daily use about five weeks later. 40 commits between 8 Oct and 15 Nov 2026.

Straight answer: It is a private local tool, not a hosted product. I don’t claim interviews, offers, time saved or a form-fill percentage, because none was measured. Recent commits carry Claude co-author trailers, and I can’t show how much code I typed myself.

  • ~1,960 automated tests verified
  • 79 employer careers boards read verified
  • 923 real sent and skipped jobs replayed to test duplicate detection verified
  • 30 of 30 jobs in a live UK search were UK or open to the UK verified
01

Overview

Job hunting eats hours of repetitive work: searching many sites, seeing the same vacancy twice, writing cover letters and retyping details into forms. This tool automates search, duplicate removal, preparation and form filling, and keeps me in control of the final submission. It is a single-user, local application, not a product for sale.

What makes it different

  • Human in the loop by design. It never presses Submit. Employer forms end with bot checks that only a person can pass, so the tool fills the form, then waits for me to tick the box and submit. Deliberately, I chose not to build CAPTCHA or bot-detection evasion.
  • A strict “never show me this again” guarantee. Any vacancy I mark Sent or Skipped, and any re-posting of it, is permanently suppressed across searches.
  • Honest handling of each source. Each job says how it can be applied for: the form can be filled, a site account is needed, or apply on their site.
  • Only sanctioned interfaces: official APIs, public feeds and employer-published job feeds. No scraping of Indeed, LinkedIn or Glassdoor.
02

How it works

From search to Sent

From search to SentSources11 platform groups,run in parallelNormalise andlink checkone Job object,usable apply linkDeduplicateids, URLs, title andcompany, similartextFilterslevel, work mode,country, age,profile rulesSuppress seenjobssent or skippednever returnScore and rankexplainable rule-based scorerDeliverthe requestednumber, deepersecond pass if shortPick: Apply orSkipone job at a time,with undoAI preparationGemini: coverletter, answers, CVemphasisForm fillingPlaywright fillsfields, attaches CVPerson submitsticks the bot check,presses SubmitSent trackingdetects submission,marks SentFrom search to SentSourcesNormaliseand linkcheckDeduplicateFiltersSuppressseen jobsScore andrankDeliverPick: Applyor SkipAIpreparationForm fillingPersonsubmitsSenttracking
  • External service
  • Service
  • Human step
  • AI model
  • Data store
Read this diagram as a list

The boxes

  • Sources: 11 platform groups, run in parallel
  • Normalise and link check: one Job object, usable apply link
  • Deduplicate: ids, URLs, title and company, similar text
  • Filters: level, work mode, country, age, profile rules
  • Suppress seen jobs: sent or skipped never return
  • Score and rank: explainable rule-based scorer
  • Deliver: the requested number, deeper second pass if short
  • Pick: Apply or Skip: one job at a time, with undo
  • AI preparation: Gemini: cover letter, answers, CV emphasis
  • Form filling: Playwright fills fields, attaches CV
  • Person submits: ticks the bot check, presses Submit
  • Sent tracking: detects submission, marks Sent

The connections

  1. Sources → Normalise and link check
  2. Normalise and link check → Deduplicate
  3. Deduplicate → Filters
  4. Filters → Suppress seen jobs
  5. Suppress seen jobs → Score and rank
  6. Score and rank → Deliver
  7. Deliver → Pick: Apply or Skip
  8. Pick: Apply or Skip → AI preparation
  9. AI preparation → Form filling
  10. Form filling → Person submits
  11. Person submits → Sent tracking
Redrawn from the pipeline in the project report. Purple steps need a person, by design.

Runs without me

  • Collecting, normalising, removing duplicates, filtering and ranking vacancies.
  • Writing cover letters and answers in a background worker, about 3 Gemini requests per job I choose.
  • Opening each shortlisted employer form and filling it in.
  • Detecting a submission to mark the job Sent, then moving to the next job.

Needs me, by design

  • Choosing Apply or Skip for each job.
  • Ticking the bot check (CAPTCHA or Turnstile).
  • Pressing Submit. The tool never does.

Sources. 79 employer careers boards read through their public job feeds (16 tagged UK, 4 Germany and 59 international), the Adzuna API (19 countries) and Reed API (UK), the NHS Jobs public feed, Workable public search, the Remotive, Jobicy, RemoteOK and Arbeitnow APIs, the We Work Remotely, Working Nomads and Jobspresso feeds, and optional Tavily web search. Himalayas is supported in code but disabled, because its apply link forces a sign-up and paid upgrade.

Duplicate prevention is a subsystem of its own. It matches by source job id, canonical URL (tracking parameters removed), title plus company plus location, and similar advert text. It collapses one vacancy listed in many places, handles adverts cut to different lengths, and lets anything I sent or skipped suppress later re-postings, including under a related employer name (“Salt” and “Salt Search”). The database has a unique constraint on the delivered job hash as the last line of defence.

A geography engine holds 51 countries with aliases, short codes and major cities. It detects a job’s country from its location and advert, classifies it as remote, hybrid or on-site, understands regions (EMEA, APAC) and restrictions (“US only”), and treats remote roles that never say where they hire from as not matching a chosen country.

03

Stack

AI models and APIs

  • Google GeminiHTTP API with a custom provider layer
  • Tavilyoptional web search, off by default

AI techniques

  • Structured JSON outputexact keys, validated, template fallback
  • Retries, failover and rate limiting80% limiter, Retry-After
  • Grounding (no invented facts)
  • Cost and token trackingper-provider usage table
  • Quality gates in codedash-removal post-processing
  • Deterministic scoringexplainable rule-based ranking

Languages

  • Pythonthe whole application
  • JavaScriptsmall inline scripts
  • HTML and CSS

Backend and APIs

  • Python standard-library HTTP serverThreadingHTTPServer, no framework
  • requests
  • PyYAML and python-dotenv
  • Threads and ThreadPoolExecutor

Frontend

  • Server-rendered HTMLinline CSS, dark theme, responsive

Data and storage

  • SQLiteWAL mode, hand-written migrations, no ORM
  • JSON and YAML configuration

Automation

  • Playwrightpersistent browser profile, Chromium or Chrome

Payments and integrations

  • Adzuna API19 countries
  • Reed APIUK
  • NHS Jobs public search feed
  • Greenhouse, Ashby, Lever, SmartRecruiters and Workable feeds79 employer boards
  • Remotive, Jobicy, RemoteOK and Arbeitnow APIs
  • We Work Remotely, Working Nomads and Jobspresso feeds

Cloud and infrastructure

  • Local Windows PC127.0.0.1:8000, not hosted

Testing and tooling

  • pytestincluding tests against a real browser
  • Claude Code
  • Git and GitHub
  • VS Code
04

AI and machine learning

Models and what they do

  • Google GeminiCover letters, drafted answers to application questions, CV emphasis suggestions and an optional AI rating of vacancies, capped per search to control cost. The default in code is gemini-3.5-flash-lite, configurable. Which model my own settings select is unverified.
  • TavilyOptional discovery of job adverts across the open web by keyword. Off by default because it spends credits.
  • Claude and OpenAINot used by the product. Claude was only the AI coding assistant that helped build it.

Techniques actually used

  • JSON-only replies with exact keys, parsed and validated. A deterministic fallback template is used when the AI fails and is recorded as ‘ai_with_fallback’.
  • Grounding rules: use only the candidate’s real background and never invent employers, dates, qualifications or achievements.
  • A privacy rule: my private notes are used only to choose jobs and must never appear in a letter.
  • A style rule that bans dash characters, enforced twice: in the prompt and by a post-processing function that also cleans text typed into forms.
  • A shared rate limiter holds Gemini to 80% of the per-minute limit and honours Retry-After on HTTP 429. It was added after real 429 errors in live use.
  • A hybrid design: cheap deterministic ranking first (title, skills with synonym groups, remote or hybrid preference, location, salary, seniority and keywords, with plain-English explanations) and AI only where it adds value.
  • Usage tracked per provider, with daily and monthly quota awareness.

Not in this project

  • Autonomous agents or tool-calling loops
  • RAG
  • Embeddings
  • Vector database
  • Speech to text or text to speech
  • Webhooks

A pipeline with human approval gates, not an autonomous agent.

05

Features

  • Step 1, Find jobs. Choose job titles, country (51), an optional city, remote, hybrid or on-site, how many vacancies, posted-within days, optional web search, optional AI rating and an optional “only jobs whose forms we can fill”. Live progress and a funnel explain why jobs were left out.
  • Step 2, Pick jobs. One job at a time with Apply (shortlist) and Skip, with undo, and a countdown of what remains.
  • Step 3, Finish applying. Start applying opens each shortlisted form in turn with details filled in, shows what to do, and marks it Sent when it detects submission. Jobs that can’t be filled are listed separately with a clear reason.
  • Sent. A count of applications sent, plus today, the last 7 days, interviews, offers and rejections, with buttons to record replies.
  • Form filling. It recognises names, email, phone, LinkedIn, portfolio, current company, notice period, salary, city and country, and handles autocomplete boxes (for example choosing “London, England, United Kingdom”). It uploads the CV and cover letter, answers some dropdowns safely, leaves consent boxes, signatures, passwords and sensitive questions alone, and lists the required fields it could not fill. Tested on live Greenhouse, Lever, Ashby and Workable forms.
  • A profile list can block named employers and sites.
06

My role

What I did. I defined the product: the search form, “never the same vacancy twice”, the Apply and Skip flow, the apply queue, the Sent counter and the platform list. I made the product and architecture decisions: a human-in-the-loop design over automatic submission, removing platforms that force sign-up, honesty about which jobs can be auto-filled, health and mental health coverage (which led to NHS Jobs), level words like “Junior” being respected, and sent or skipped jobs never returning. I supplied the domain rules in my profile (target roles, excluded roles and industries, salary floor, blocked employers). I ran the product against real job sites daily and reported precise defects, and much of the accuracy work came from those reports.

What Claude Code did. All the code, tests and tooling, working interactively under my direction. Commits are authored under my git identity and recent ones carry “Co-Authored-By: Claude” trailers, so I can’t verify how much code I typed myself.

07

Engineering challenges

API rate limits

Gemini returned HTTP 429 errors in live use.

How it was solvedHide

Fixed with a shared rate limiter at 80% of the per-minute cap, Retry-After handling and fewer requests per job.

Duplicate vacancies

Platforms list one job many times, and I once applied twice.

How it was solvedHide

Adzuna sent the same advert cut at about 400 and 500 characters, so similarity fell below the threshold. I added containment matching, related-employer matching for agencies that re-post under names like “Salt” and “Salt Search”, and made skipped jobs suppress re-postings. Checked by replaying the real history of 923 sent and skipped jobs.

Wrong-country results

A UK search returned a California job because its advert said “global”, and a US “Distributed” role.

How it was solvedHide

The country rules were tightened. Verified live: 30 of 30 jobs in a UK search were UK or open to the UK, with 115 foreign jobs filtered out.

Seniority

“Junior AI” returned senior roles.

How it was solvedHide

Level words in a search became a hard requirement (junior, senior, lead), with synonyms (jr, graduate, entry level). Verified: 3 of 3 results were junior.

Autocomplete form fields

Location boxes accepted only a chosen suggestion, so typed text was discarded.

How it was solvedHide

The filler now selects a suggestion, preferring my own country over a same-named town elsewhere (“London, England”, not “London, Ontario”), reads labels through aria-labelledby and detects already-chosen values.

Slow sites

The form filler gave up after 20 seconds on slow pages.

How it was solvedHide

It now waits 45 seconds and retries once.

Browser threading

Playwright’s sync API is bound to the thread that created it.

How it was solvedHide

The apply queue runs in a dedicated thread that owns the browser.

Honest scope

Employer forms end in bot checks.

How it was solvedHide

Instead of building CAPTCHA evasion, I built the human-in-the-loop queue, so the person only ticks the bot check and presses Submit.

Missing coverage

A search for mental health found nothing, because no platform here carried NHS posts.

How it was solvedHide

I added the NHS Jobs public feed (a live London search then returned 30 mental health roles), fixed an “adult” industry exclusion that wrongly removed “Adult Mental Health” jobs, and stopped “Physical Health Nurse” matching “Mental health”.

Text hygiene

Generated text must never contain dash characters.

How it was solvedHide

Enforced with both a prompt rule and a post-processing function.

08

Status

In active personal use, running locally on my Windows PC. There is no public deployment and no demo link. A typical search takes 30 to 60 seconds in live tests, with sources running in parallel.

Complete

  • Search across all sources, filtering, ranking and duplicate removal.
  • Apply and Skip review, AI preparation, form filling, the apply queue and Sent tracking.
  • Employer blocking, NHS Jobs search and Workable search.

Partial or still open

  • Form filling is partial by nature. Employer forms differ and always include questions or bot checks that need a person. My samples filled 5 to 10 fields per form.
  • Adzuna, Reed, NHS Jobs and the remote boards open only the advert, and the application is made on the employer’s own site.
  • Workday and Taleo sites need an account and can’t be filled.
  • The latest work is on the vacancy-search branch. A merge into main is not confirmed.

Not built, and not claimed

  • Autonomous agents, RAG and voice.
  • Multi-user support and cloud hosting.
  • Automatic submission, which is intentionally excluded.
09

Evidence

  • ~1,960 automated tests 1,964 passed on the last full count I saw, with a few added since. About 16,100 lines of tests against 20,700 lines of application code, including tests against a real browser.verified
  • 79 employer careers boards read 54 Greenhouse, 16 Ashby, 4 Lever, 4 SmartRecruiters and 1 Workable.verified
  • 923 real sent and skipped jobs replayed to test duplicate detection verified
  • 30 of 30 jobs in a live UK search were UK or open to the UK With 115 foreign jobs filtered out.verified
  • 51 countries in the geography registry verified
  • 11 platform groups searched verified
  • 30 to 60 s for a typical search in live tests verified
  • 622 vacancies marked sent in the database My real usage: 1,212 jobs stored, 37 searches run and 1,089 vacancies delivered.verified

Code worth reading (private tool)

  • app/search/engine.pyThe pipeline and level-aware title matching.
  • app/processing/geography.pyCountry and work-mode detection across 51 countries.
  • app/processing/deduplicator.py and app/search/store.pyDuplicate and never-show-again logic.
  • app/applications/autofill.pyForm filling and autocomplete handling.
  • app/applications/session.pyThe apply queue and submission detection.
  • app/applications/preparation.pyPrompt design for letters and answers.
  • app/providers/throttle.pyRate limiting.
  • app/processing/scorer.pyExplainable scoring.

What I can't verify, so I don't claim

  • Interview, offer or hiring outcomes. The Sent page showed zero interviews and offers when I last looked.
  • Time saved, and any form-fill percentage. Neither was measured.
  • Whether the code repository is public, so it is not linked.
  • Whether a daily scheduled search task is registered on my PC.

verified checked against the code, tests, logs or the live system. reported from my own notes or hands-on use. unverified could not be checked, so it is not claimed.