Job Search Bot

AI Coding + QA Testing = Now & Future.

(AI Coding = building AI agents, chatbots and websites)

I work hands-on with agentic AI development tools that make it possible to specify, build and test complete, production-ready solutions quickly and efficiently — without having to master traditional programming languages such as Python or PHP. Throughout the project, my role has been to specify precisely what should be built, test the output thoroughly, and find concrete bugs and errors in the AI's output and logic before anything went into production.

How it works

Three steps from search term to a finished application.

Search for jobs

The bot fetches jobs directly from Jobindex.dk and Jobnet.dk based on your search terms, with an option to filter by city. Results from several sources are merged, and duplicates are removed automatically.

AI compares with your CV

Each job posting and your CV are sent to Google Gemini, which assesses the real professional overlap — not just shared words — and returns a match percentage with a short explanation.

Get finished documents

Pick a job, and the bot drafts an application in Danish plus the job posting as a PDF with the contact details extracted. You review and send it yourself.

Screenshot: the misspelled city name 'Arhus' is automatically recognised as Aarhus — A misspelled city name is recognised automatically against the official list of Danish municipalities — a reliable lookup, not AI guesswork.

Agentic search: the AI thinks for itself when a search comes back empty

The bot has no fixed, predefined list of synonyms. Instead, the AI decides for itself which similar word to try when a search returns 0 results.

0 results

The search term doesn't appear directly in any job posting — the bot automatically brings in the AI instead of just showing "nothing found".

if the word resembles a real word

The AI chooses for itself

Gemini suggests one broader or synonymous Danish word, runs a real search with it, and looks at the actual result — at most 2 attempts, never more.

if the word is nonsense

Honest refusal of nonsense

If the search term makes no sense, the AI says so honestly instead of guessing a random word.

🤖

Real example — search term with a typo: "boghlder" (Danish for bookkeeper is "bogholder") The word 'boghlder' is a typical misspelling of the Danish job title bogholder (bookkeeper), so I'll try searching with the correct spelling instead. → Tried «bogholder»: 61 jobs found.

Screenshot: the search term 'boghlder' is automatically corrected to 'bogholder' by the AI, 61 jobs found — A typo in the search term ("boghlder") is corrected by the AI to the right Danish word — 61 jobs found straight away.

🤖

Real example — search term: "abrakadabra" "abrakadabra" is a magic word and meaningless text in a job context, and so it cannot be interpreted as a real field or a job title.

Screenshot: the search term 'abrakadabra' is typed directly after «What now?» instead of pressing Show the list again — You can type a new search term directly instead of pressing «Show the list again».

Screenshot: the bot honestly rejects 'abrakadabra' instead of guessing a random word — The bot honestly refuses instead of searching for a random word.

Honest match assessment

It isn't just a search for matching words. If the CV has no relevant experience for a position, the bot honestly shows 0% — no "consolation percentages". The job still stays on the list, because whether to apply is always the person's own choice.

Not just a keyword match, as you often see on other platforms — the AI assesses the real substance: does the experience actually match the role's requirements, or is it only a superficial overlap of words? If something real is missing, the bot says so honestly — instead of showing a nice but misleading percentage.

Example of a positive match: bookkeeper CV against a Financial Bookkeeper position, 90% — **Good match** — bookkeeper CV against "Finansbogholder" (Financial Bookkeeper): 90%

Example of a partial match: bookkeeper CV against a Controller position, 25% — **Partial match** — bookkeeper CV against "Controller til CISU": 25%, with an explanation of why the role requires more than accounting experience

Example of no match: bookkeeper CV against a painter and decorator position, 0% — **No match** — the same bookkeeper CV against "Bygningsmaler" (Painter and decorator): 0%, no consolation percentages

What it generates for each job

When you pick a specific job, the bot drafts 2 documents in under a minute:

Application in Danish Based on real facts from the CV, with no invented experience. If something is missing, it is mentioned honestly but tactfully, as a willingness to learn it quickly.
The job posting as a PDF The job posting in Danish with automatically extracted contact details — name, phone and email — so you have everything in one place.

Afterwards, the application can be checked and corrected with two buttons:

❓ Verify: check for invented facts A new, independent AI review compares the application with the real CV and flags every claim that isn't in the CV. At the same time, it checks whether the application completely misses an important requirement from the job posting, either as experience or as an honestly mentioned gap.
🔧 Fix based on Verify If Verify has found something, the bot corrects the application in a targeted way, only in the flagged places, and recommends running Verify again on the corrected version.

The bot never sends anything by itself. It only drafts — the final step (review, adjust, send) is always the person's own.

Verify — the AI checks itself before you do

With the "❓ Verify" button, the AI compares the drafted application again — this time against your original CV and the job posting itself. If it finds a claim that can't be confirmed, or a requirement that isn't addressed at all, you can fix it with one click ("🔧 Fix based on Verify") and check again until everything is green.

Verify result: found a claim not supported by the CV, and a requirement that isn't addressed — **Verify finds something** — a claim with no support in the CV, and a requirement from the posting that isn't mentioned

Corrected application delivered as a PDF after pressing Fix based on Verify — **Fixed with one click** — only the points found are changed, the rest of the application is left untouched

New Verify check: everything confirmed, check passed — **Clean check** — everything confirmed by the CV, all requirements addressed

Download Verify documentation (PDF, Danish)

Quality assurance in practice

Not a theoretical QA checklist — concrete bugs, found through repeated testing with real data during development, and fixed right away.

A misleading "nothing found" message The same text was shown whether the search actually returned 0 results or the results had simply been shown before — now two precise, different messages.
The «Search term:» line silently disappeared When the AI corrected a typo in the search term, the line showing which word matched disappeared — because the bot was still comparing against the original, misspelled word.
A misspelled word silently dropped With several search terms at once, a misspelled word disappeared completely, with no correction or message, if just one of the other words already returned results.
A promise the bot couldn't keep The footer suggested that more results could be found by refining the search — but in reality they were just lower-scored and dropped for good.
A hanging AI call with no error message A Gemini call without a timeout could, in rare cases, stop the bot completely, without a single error in the log. Solved with an explicit 45-second limit.
Data that disappeared with every update The database (CV, search history) wasn't on a persistent disk and was reset with every code update — fixed, together with automatic deletion of inactive users' data after 30 days for privacy reasons.
Management positions scored too high A Product Manager position got an 80% match because the AI saw the shared topic "AI" — but the CV showed a purely technical background with no product management experience. The job function itself (leading vs. building) wasn't assessed as a separate factor. Fixed with an explicit example in the assessment prompt.
The explanation switched language without warning The instruction to the AI was itself written in a different language from the one it had to answer in — which in rare cases "leaked" into the answer itself. Fixed by making the instruction language consistent.
Verify raised a false alarm An honestly mentioned gap in the application ("no experience with X yet, but a fast learner") was flagged as an invented claim. Fixed: an honest gap is not a factual error — only claims with no support in the CV are.
Verify explained in English When the CV or the job posting was in English, Verify's explanations also came in English, even though the rest of the bot is in Danish. Fixed: the explanations are now always in Danish.
Contact details that were right there got missed A job ad clearly named a contact person with phone and email near the bottom of the text — but the extraction returned "not stated" for all three fields. Not a character limit, just the AI not reading the posting carefully enough to the end. Fixed with a clearer instruction to read the whole text.

Technical

Built as a standalone Python project with a focus on reliability and simple operation.

Python python-telegram-bot Google Gemini API Gemini function calling (agentic search) SQLite fpdf2 (PDF generation) Jobindex.dk + Jobnet.dk Hosted on Railway

Specified, built and tested with agentic AI development tools — without having to master these languages in the traditional sense.

Not just a prototype: a 45-second timeout on every AI call (so a hanging call never stops the bot without warning), a persistent database that survives every new deployment, automatic deletion of inactive users' data after 30 days for privacy, and a usage limit on the public demo to keep costs under control.

See the bot in action