AI projects and programs

Eight LLMs put to the test

Scalability of semantic intelligence in LLMs: a cross-tier evaluation of zero-shot detection of morphosyntactic anomalies (paper in Italian). By Giuseppe Testa, Edoardo Pio Di Muro, Giovanni Di Domenico, Francesco Riemma.

My role: Software development (generators and analysers in Python), data extraction, tests on Claude, manual validation of ambiguous cases, writing of the paper in CEUR-WS format.

The method

• 12 real Italian texts from the Paisà corpus, about 1,200 words each.
• 10 – 12 neologisms inserted per text with spaCy (Garlio, Vembrare, Sedialoso), at most one per sentence.
• The same validation prompt for every model, in zero-shot mode.
• Automatic comparison (exact and lemma match) plus human validation of doubtful cases; calculation of precision, recall and F1.

F1 score by model

• Gemini 3.1 Pro (Pro)84.4%
• Gemini 3.1 Flash (Free)81.4%
• ChatGPT o3 (Pro)77.1%
• Claude 4.7 Opus (Pro)70.1%
• Claude 4.6 Sonnet (Free)69.5%
• ChatGPT 5.3 (Free)67.6%
• Copilot Smart (Free)47.8%
• Copilot Think Deeper (Pro)42.5%

Slides, banners and screenshots are shown as in the original projects, in Italian.

More programs born from an idea

Each one starts from a concrete problem. The code stays private: here I explain the idea, how I developed it by guiding AI and what it does, with real screenshots.

Video Saver, Android app

Save a public video in two taps

The problem was simple: save a video seen on social media without going through websites full of ads. I imagined an app with a single field, “Paste the link”, that could also be used from the Share button of any app.

Then came the question every user asks: which social networks does it really work with? So I added a compatibility test that tries the preview of public links from each platform, without downloading anything, and saves a report.

The role of AI: I defined the features, interface and test cases; the code was written with AI assistants under my direction, then tested on the phone.

Result: Latest test: 20 social networks work, 2 only when logged in, 4 do not work. The app only downloads public content.

SPC, a football prediction system

Serie A 2026/27, from statistics to a readable result

I wanted to find out whether a serious statistical model could give predictions that make sense even to people who are not fluent in probability. SPC uses the Dixon-Coles model with a Bayesian prior: it estimates each team’s expected goals and derives every possible score from them.

The real challenge was communication: showing a recommended score with its probability, the two closest alternatives and the derived markets (1X2, under/over, both teams to score) without hiding the uncertainty. One option asks AI to factor in news and line-ups.

The role of AI: I chose the model, wrote the requirements and designed the screens (Prediction, Teams and form, Data sources, Validation); development was done with AI and checked on the Validation tab.

Result: Example: Atalanta–Bologna, 1–1 recommended with a 13.8% probability, expected goals 1.12 and 1.09, draw at 30%.

VDisk Converter

Move a virtual machine without losing a single bit

Anyone who works with virtual machines knows the problem: every program wants its own disk format. I designed a converter that automatically recognises VHDX, VHD, VMDK, VDI, QCOW2 and IMG/RAW and converts them, for example, to VMDK for VMware.

The interface shows the conversion as an arrow from one format to the other. You can convert a single disk or an entire folder, and at the end the program checks that the copy is identical to the original.

The role of AI: I defined the formats, options and safety checks; the code was produced with AI assistants and I tested it on real disks.

Result: Version 0.2: 6 formats recognised, copy verification, up to 16 parallel streams.

Semantic CV system

The project you are looking at

This website, a CV builder with 7 roles, DOCX/PDF export and a QR code, and a test environment in which 100 virtual HR recruiters assess the CV and the site at every version.

The role of AI: Designed with AI as a working tool, from the data structure to the tests.