I am a chartered accountant. I also build the systems that do my profession's work.

Claim-verification service for insolvency trustees

A court-appointed trustee uploads a creditor's claim as a PDF. The service identifies the creditor, shows the price before it processes anything, and returns a verification sheet in Word and a table in Excel, ready for the case file once the trustee has reviewed them. I hold these appointments myself. This is my own profession's work, automated by someone who has done it by hand under professional liability, and it is now used by other trustees on real filings.

• No external AI model is ever called. Extraction is deterministic and runs locally. That is a risk decision, not a technical limit: insolvency case files must not leave the machine to be processed by a third party.
• Reconciled code by code against the tax authority's own ledger extract. Where a figure in the claim does not match, the row is flagged rather than silently accepted. OCR handles scanned claims and files whose text layer is corrupted.
• Priced per document, charged only on success. The first check is free at any complexity, then EUR 2.99 per claim, no subscription. A failed job bills nothing.
• Adversarially reviewed before release: a written red-team report with a pre-mortem and a criticality matrix naming who would challenge each weakness, from the data-protection authority to the supervising judge. API keys are stored only as SHA-256 and shown once; temporary files are cleared at start-up and in the failure path.
• The privacy notice is enforced by the code, not merely published. Visit records are deleted after 90 days and registration IP addresses after 30, on every start-up. A retention period a user can read is a promise, and a promise belongs in the software rather than in a document nobody checks.
• The system never decides. Its output is a draft the trustee reviews, completes and signs.

Open the service Read why I built it Both in Italian

Public fraud lookup service

One question, one answer: is this phone number, website or email address a scam? It searches a database assembled from dozens of open feeds and returns a verdict. Free and with no registration. You can check the claims below by using it.

• A single SQLite database holding 4,340,722 records in WAL mode, the count published on the home page, with schema and indexes created and migrated on first access.
• Plain PHP 8 with no framework and no dependencies, chosen deliberately so it runs reliably on shared hosting.
• Deploy discipline: nothing overwrites production before the live file is downloaded and compared, and every release is verified by behaviour on the real site.

Try it Read why an accountant built it Service in 6 languages, article in Italian

Autonomous retail pricing pipeline

You paste a shopping list in plain language. It tells you, line by line, the cheapest matching offer in the valid flyers near you. The hard part is not the search, it is that the source documents are image-only PDFs with no text inside.

• Runs unattended. The nightly cycle starts at 02:00 and finished ~1,500 offers across six retail chains in 60 minutes with no human involved.
• Extraction uses vision models with JSON-schema output. Measured accuracy on the test set: 100% across ~42 products, including dense pages and non-food items.
• It refuses to publish what it cannot verify. Unit prices are recomputed and compared against the price-per-kilo already printed on the flyer, handling packaging formats and multipacks. Anything that does not reconcile is quarantined for review.
• Model failover proven live: when one vision provider rejected pages, a second recovered them. One generic ingestion framework with an adapter per chain, not single-use scrapers.

Real-time market reaction platform

A subscription product and a paid public API. It watches for public posts, measures how the monitored markets moved in the five minutes after each one, issues a verdict, and delivers the event to subscribers and to third-party clients.

• The whole chain is automatic: detect, deduplicate, filter, activate, measure at T+5, close on recovery, judge, notify, distribute. Seven production cron jobs plus a wake-up channel.
• Three market-data providers with controlled degradation. When the metered provider exhausts its quota the system falls back, and every measurement records which source produced it.
• A symbol is not accepted because it responds, but because it responds for the same instrument. Cross-source verification catches the ones that quietly do not.
• I designed the commercial side too: subscription tiers, an annual API plan with key and rate limits and three webhook endpoints, a delayed sandbox tier reusing the same code path, and contract-based redistribution.

Guardrailed AI for regulated professionals

An agent platform for lawyers and tax advisors, built around one non-negotiable rule: accountability stays with the human professional. Every output is a working draft subject to their verification, and the product is engineered so it cannot pretend otherwise.

• Seven absolute prohibitions encoded as guardrails: never invent legislation or case law, never cite a source absent from the provided context, never guarantee an outcome, never present a deadline as final, never fill a gap with an assumption dressed as fact.
• Every agent must add at least two guardrails specific to its own subject matter, and an automated test fails the build if it does not.
• A fixed ten-section output contract, including missing data, stated assumptions, risks, and a verification checklist for the professional.
• Three-label source attribution separating uploaded documents, the knowledge base, and everything still to be verified.

Multi-provider market data API

A financial data API behind one interface, running on a hardened VPS. Honest status: it is in service and healthy, and it does not yet have external consumers.

• Four providers behind a single contract, with credit rationing driven by a market clock: closed markets consume no quota, because wasted calls are wasted money.
• Rootless Docker, bound to loopback only, never exposed directly to the internet.
• Client keys are stored only as SHA-256 hashes. No secret ever enters version control.

Trilingual UAE pre-engagement platform

A platform that structures and qualifies European founders' UAE enquiries before they reach a licensed local operator. English, Italian and Arabic, with right-to-left support.

• Compliance rules are enforced in the product, not in a policy document: no outcome promises, no fabricated metrics or testimonials, no company logos presented as clients.
• Contact details are disclosed only through an explicit consent flow, and in demonstration mode they do not exist in the data at all.
• The scope of what the service does not do is stated openly: it does not incorporate companies, issue licences, file visa applications or give advice.