SIMONE PERINI @TZII

AI & software engineer · Italy

Hi, I’m Simone.I like makingthings work.

On GitHub, I’m tzii. I build AI systems, work on apps, and follow an idea until there’s something you can actually try.

Games, AI & the occasional rabbit hole.

All my work GitHub

01 / Selected work

Things I’ve put time into.

Five projects. Different problems.
The decisions behind each one.

Evidence-first document intelligence
Architecture sketchPython / FastAPI / Qdrant

01 / AI & dataPublic case study

italianRAG

Answers you can trace.

Italian corporate PDFs, typed extraction, and retrieval with source evidence. An internship project built around a simple rule: show where the answer came from.

Explore case study

Document intelligence · Internship project

italianRAG

An answer is only as good as its evidence.

A document-intelligence workspace for Italian corporate PDFs. Developed during a three-month AI & Data internship at InfoCamere, it turns financial statements, statutes, and registry reports into typed facts and source-linked answers.

PythonFastAPIPydanticSQLiteQdrant

The problem

Corporate documents do not share one clean layout. A value can be missing, ambiguous, or contradicted by another document. A fluent answer is not enough when someone needs to check the original evidence.

What I worked on

I worked on the FastAPI service and local workspace, document-specific extraction, typed field states, SQLite fact modelling, Qdrant retrieval, metadata filters, and reranking experiments. Cross-document routes compare financial years and reconcile statute and registry information.

The decision that matters

Deterministic routes handle supported structured fields. Retrieval handles broader questions. Both keep page references and processing traces close to the result. The default parser stays lightweight; OCR and vision-model recovery are opt-in rather than a hidden dependency.

Scope, not a universal accuracy claim

The public case study records 116/116 answer-correct, citation-supported, and gold-verified cases on its defined deterministic benchmark. This does not establish 100% accuracy on arbitrary PDFs, unseen companies, OCR, or open-ended RAG. Weak or conflicting evidence still needs human review.

Private implementation. Public engineering story.
Pefforza 4

Independent browser mini-demo6 × 7

02 / Game AI & visionOpen source

Pefforza 4

A game with a lot going on.

Connect Four with physical-board vision, bitboard search, an exact solver, and reinforcement learning. A small game that leaves very little room for hand-waving.

Explore case study

Game AI · Computer vision

Pefforza 4

A familiar game. A surprisingly rich engineering problem.

A Connect Four project that connects computer vision, classical search, reinforcement learning, and a usable game interface. Play in the terminal or Pygame, or point a webcam at a physical board.

PythonOpenCVGymnasiumPPOBitboards

What it does

Webcam calibration and colour classification reconstruct a real 6 × 7 board. State validation checks gravity, turn counts, legal transitions, and terminal positions before the AI is allowed to recommend a move. Optional voice commentary fails softly when audio is unavailable.

How the AI is structured

The project separates random and heuristic opponents, depth-limited bitboard alpha-beta search, a bundled PPO policy, and a time-budgeted exact solver. Evaluation alternates sides rather than letting first-player advantage quietly skew the result.

Correctness under a time budget

The exact solver uses threat pruning, symmetry-aware caching, and packed transposition entries. A result distinguishes a proven outcome from a fallback. On timeout, the solver must not return a root move it has already proved losing when an unresolved alternative remains.

What the browser demo is, and is not

The little board on this portfolio is an independent JavaScript interaction. Its opponent takes an immediate win, blocks a threat, and prefers central safe moves. It does not run the Python engine, PPO checkpoint, webcam pipeline, or exact solver. The original impossible tier only guarantees a proven move when the proof fits its budget.

Python · OpenCV · PPO
KeeFetch
Ranked, not just downloaded.
Favicon retrieval for KeePassConcept illustration

03 / Desktop toolingOpen source

KeeFetch

The little things, done properly.

A concurrent favicon downloader for KeePass. Candidate ranking, explicit privacy choices, and release checks turn an everyday annoyance into a maintained tool.

Explore case study

Desktop tooling · Maintained project

KeeFetch

Small icons. A lot of decisions behind them.

A KeePass 2.x plugin that retrieves favicons without making the application wait on one website at a time. The interesting work is in candidate quality, concurrency, privacy trade-offs, and release reliability.

C#.NET FrameworkKeePassConcurrencyTesting

What it does

KeeFetch fetches icons concurrently, examines site icons and manifests, ranks provider candidates, and deduplicates image data with SHA-256. It also handles KeePass reference URLs and Android application identifiers.

Selection, rather than first response wins

An available image is not necessarily the right image. The selector ranks site-backed and resolver-backed candidates above synthetic fallbacks. The current source includes study-selected Fast, Balanced, Privacy, and Precise profiles, with a Custom profile for explicit control.

Engineering beyond the download

Recent merged work scopes TLS and certificate handling to the plugin’s HTTP client instead of changing the whole KeePass process. Release checks cover versions, the tracked PLGX source manifest, artifact checksums, and permissions. Source capabilities and published downloads are documented separately.

The privacy boundary

No telemetry does not mean no external requests. Resolver services can receive entry domains. The Privacy profile disables configured resolvers, but site-linked assets and redirects may use other hosts. Diagnostics can contain entry titles and URLs. Private-host filtering is not a complete DNS or pre-request network policy.

C# · .NET · KeePass
ThystTV
ThystTV full-screen Android player with floating chat
Video first. Chat close by.
App screenshot / Xtra-based forkScreenshot unavailable / illustrationAndroid

04 / AndroidXtra-based fork

ThystTV

More watching. Less fiddling.

An Android Twitch client focused on playback controls, floating chat, local stats, and large-screen comfort. Built on Xtra, with credit where it belongs.

Explore case study

Android · Open-source fork

ThystTV

Keep the video first. Make the rest work around it.

An Android Twitch client based on Xtra. My fork focuses on playback controls, floating chat, local viewing statistics, large-screen layouts, and the maintenance that keeps a media client usable.

KotlinAndroidRoomCoroutinesRelease engineering

The viewing experience

Gesture-based seeking, playback speed, brightness, and volume controls reduce trips through menus. Floating chat keeps the conversation visible during full-screen playback. Local statistics describe viewing patterns without needing a separate analytics service.

The work behind the interface

Maintaining the app also means keeping upstream changes reviewable, handling network recovery, and avoiding brittle release steps. Recent bookmark-refresh work treats metadata fetch and database application as one retryable operation while preserving coroutine cancellation.

Trustworthy release artifacts

The repository documents APK checksum and signing-certificate verification. Release work checks structured signer data and promotes verified candidate bytes rather than silently rebuilding a different artifact. Device testing and release-specific checks remain distinct from a successful local build.

Credit and boundaries

ThystTV is a fork of Xtra, not a client built entirely from scratch. Xtra deserves credit for the base client and architecture. ThystTV is independent of Twitch and Amazon. The project screenshots show the Android app; this portfolio is not an embedded streaming client.

Kotlin · Android · Room
ScrapingStore
ScrapingStore generated product-data dashboard
scrape → clean → something useful
Project screenshot / generated snapshotScreenshot unavailable / illustrationPython

05 / Data engineeringStudent project

ScrapingStore

Data, all the way through.

Browser scraping, validated records, SQLite storage, and dashboards you can actually inspect. A complete sandbox pipeline, with its limitations left visible.

Explore case study

Data engineering · Student project

ScrapingStore

The data pipeline does not end at extraction.

An end-to-end Python project for collecting sandbox product data, cleaning it, storing it, and making it inspectable through dashboards and Power BI-ready exports.

PythonPlaywrightPandasSQLModelJinja2

The full route

Browser automation extracts product data from the Oxylabs e-commerce sandbox. Pandas normalizes and deduplicates the records. SQLModel and Pydantic define the storage contract, SQLite persists it, and Jinja templates generate a dashboard. CSV export makes the same data available outside the web interface.

Different tools for different pages

The project includes both a static BeautifulSoup scraper and a Playwright browser scraper. The target sandbox is JavaScript-rendered, so the browser path is required for that target. The static implementation demonstrates a separate pattern rather than pretending it can execute JavaScript.

Making the result inspectable

A Typer and Rich command-line interface exposes scraping, export, and report generation. Modern and terminal-style dashboards present the output. Tests, Docker configuration, and GitHub Actions support repeatable development.

Known limits

Selectors are sandbox-specific. Products are upserted by source identity. Records with weak or legacy identities are not suitable for reliable price-history comparisons. The public dashboard is a generated snapshot, not a promise of live inventory. Adapting the pipeline to another source requires deliberate selector and data-identity work.

Playwright · Pandas · SQL

02 / The person behind the projects

A bit more
about Simone.

Pefforza is a good introduction to how I work: start with something familiar, then keep asking what else it could do. A game becomes a search problem, a learning environment, and a reason to point a webcam at a real board.

My focus is applied AI: document intelligence, Python backends, retrieval, and evaluation. But I also build desktop tools, work on Android apps, and use games to explore how intelligent systems behave. I like getting the small details right too: a move the AI can explain, a retry that works, or an icon that finally looks right.

Education
Higher Technical Diploma, AI & Data
EQF level 5 · Final grade 107/110
Experience
AI & Data internship at InfoCamere
Document-intelligence engineering
Core work
Python, FastAPI, typed APIs
Retrieval, extraction, evaluation
Also on the workbench
Kotlin / Android, C# / .NET
Computer vision, game AI, tooling
01

Keep the evidence close.

A source-linked answer. A reproducible benchmark. A solver that tells you when it has a proof. Confidence is useful when you can inspect what it rests on.

02

Finish the unglamorous parts.

Package checks, retry behaviour, keyboard controls, and honest release notes. The details that make a project useful after the screenshot has been taken.

03 / Get in touch

Let’s talk
about what’s next.

Connect on LinkedIn Find me as tzii
Behind the build / Engineering notes

↑ ↓ to explore · Enter to openEsc to close