
Data and AI work-study observatory
A pipeline that collects every work-study offer in France, has them analyzed by AI and reveals the most in-demand technologies for each role in a Power BI dashboard.
Completed
October 2026
Personal
Data / AI
Technologies
01The project
When looking for a work-study position in data or AI, one question keeps coming back. Which technologies do you really need to master? Python, SQL, Power BI, the cloud? Rather than relying on generic advice, I wanted to find the answer directly in the job offers.
So I built an observatory that collects every work-study offer published in France, isolates the ones related to data and AI, has them analyzed by a language model, and presents the most in-demand technologies for each role in a Power BI dashboard.
Data and AI work-study observatory
Offers collected from October 21 to 26, 2026, source La bonne alternance
Work-study offers in France
12,841
Data work-study offers kept
35
Duplicates detected
7
Work-study offers by role
Most requested technologies
- Power BI31%
- Python31%
- SQL29%
- Excel26%
- Git17%
- Pack Office14%
- GCP11%
- DAX9%
- Docker9%
- Tableau9%
Only technologies mentioned in at least 2 offers are shown.
02Hundreds of thousands of rows to sort
The offers come from La bonne alternance, the French public service that centralizes work-study offers. Its API provides a full export of about 570 MB, which I download as a stream and read offer by offer, without ever loading the whole file into memory.
The first finding was surprising. Out of the 295,000 rows in the export, 96% are not real offers but companies suggested for spontaneous applications. Only about 12,800 real offers remain, and only around sixty of them are related to data and AI.
To spot them, a pre-filter combines keywords in the title with ROME codes, the job classification of France Travail, the French public employment service. It deliberately casts a wide net, even if it lets some false positives through, which the AI then removes.
03Spotting offers published several times
The same offer is often posted on several websites, with a different ID each time. Without a fix, an offer published three times would have counted three times in the statistics.
So I built a deduplication step that splits each description into small five-word chunks and measures how much two descriptions overlap. Along the way, I discovered that a typographic apostrophe or a "œ" was enough to make two identical copies look like different offers. I had to normalize the texts before comparing them.
04AI to read the offers
Each offer is then sent to Claude Haiku, which returns a structured answer. Is it really a data role? Which job? Which technologies are mentioned? Is it a real work-study contract, or an internship in disguise?
Most of the work went into the prompt. A first version counted as data roles some finance or safety jobs that simply involved a few dashboards. Worse, two identical descriptions had received two different answers. I tightened the rules, with one clear principle, when in doubt, it's not a data role. I also versioned the prompt, so that only offers never analyzed with the current version are sent to the model. Analyzing a day of offers costs a few cents.
05What the data says
Among the offers collected from October 21 to 26, 2026, 35 data and AI work-study offers were kept. Power BI and Python are requested in 31% of offers, SQL in 29%, Excel in 26%.
The breakdown by role is the most telling. For a data analyst or BI position, Excel and Power BI are requested in half of the offers, far ahead of Python. For data engineering, development tools dominate, with Git, Docker and the cloud. And for AI engineering, the tools mentioned are very scattered, a sign of a young field without a standard toolkit yet.
06What I take away from it
This project taught me that raw data often lies. Recruiters disguised as offers, offers published three times, misassigned job codes, internships presented as work-study contracts. Most of the work wasn't the analysis, it was the cleaning.
I also learned to keep a critical eye on my own figures. Adding just five offers moved SQL from fifth to third place. With 35 offers, my results show a trend, not the truth, and I'd rather say so clearly. That's why the dashboard only shows a technology if it appears in at least two offers, and only details a role from five offers onward.
Finally, working with a language model inside a data pipeline showed me that AI is an excellent sorting tool, as long as it gets precise rules and its answers are checked. And this project gave me a very concrete answer for my own search, since Power BI is among the most in-demand skills.
The next step is to automate the collection every day, to accumulate offers week after week and make the figures more and more reliable.
07Tech stack
Python for extraction, cleaning and statistics, the La bonne alternance API for job offers, Claude Haiku through the Anthropic API for analyzing the offers, and Power BI with DAX for the dashboard.