Matteo GalloSoftware & AI Engineer · LausanneAvailable

I build AI products you can verify.

Complete products, from the interface to production. Every measure is published, defects included.

Live demo: Verity, my RAG agent, on a test contract

Service Level Agreement

This Service Level Agreement ("SLA") applies to the services provided by Acme Corporation under the Master Services Agreement.

1. Uptime Commitment

Provider commits to a monthly uptime of 99.9% for the production service. Uptime is measured per calendar month and excludes scheduled maintenance windows.

2. Service Credits

If monthly uptime falls below the commitment, the Customer is entitled to a service credit calculated as a percentage of that month's fees:

Uptime from 99.0% up to 99.9%: 10% credit.
Uptime from 95.0% up to 99.0%: 25% credit.
Uptime below 95.0%: 50% credit.

Service credits are the Customer's sole and exclusive remedy for availability failures and must be claimed in writing within thirty (30) days of the affected month.

3. Support Response Times

Provider responds to support requests according to severity:

Priority 1 (critical, service unavailable): within one (1) hour.
Priority 2 (high, major feature impaired): within four (4) business hours.
Priority 3 (normal): within one (1) business day.

4. Scheduled Maintenance

Routine maintenance is performed on Sundays between 02:00 and 06:00 UTC. Provider gives at least forty-eight (48) hours' advance notice of any scheduled maintenance, which does not count against the uptime commitment.

1 The question

What is the response time for a Priority 1 support request?

3 The answer, verified

For a standard Priority 1 (critical, service unavailable) support request, the response time is within one (1) hour. If the customer has Premium Support, the Priority 1 response time is shortened to thirty (30) minutes, available 24 hours a day, seven days a week.

Cited: SLA contract (section 3), Northwind order form (section 3)

Quote found word for word in the document
3/3 claims supported by their quote, according to the judge
confidence 0.99

1 The question

What is the response time for a Priority 4 support request?

3 The answer: a refusal, verified

I don't know based on the provided documents.

Refused by the model, confidence 0.97

Closest passage: SLA contract (section 3), judged insufficient

The SLA section (document index 0) explicitly lists only Priority 1, 2, and 3; no Priority 4 tier exists in any evidence, so the answer correctly refuses to fabricate a response time.

Run source

/* Without the script the tabs cannot switch: show both cards and no inert controls. */ [data-verity-demo] .tabs { visibility: hidden; } [data-verity-demo] .cards { display: block; } #verity-panel-refusal[data-off] { visibility: visible; opacity: 1; transform: none; margin-top: 1rem; } /* Nothing scrolls the window to the passage: the whole contract is shown, the cards below it. */ [data-verity-demo] [data-sheet] { height: auto; padding-bottom: 1.25rem; mask-image: none; } [data-verity-demo] .cards { margin: 1rem 0 0; }

0/31answers with a hallucination, according to an LLM judge, among the answerable questions Verity answered

17/20unanswerable questions correctly refused

3/20unanswerable questions it answered anyway: its main defect

+ 01

Projects

No. 01

Verity

A RAG agent over company documents: it answers by quoting word for word, refuses when its evidence is weak, and publishes how often it refuses rightly.

17/22 refusal precision
κ 0.71 judge agreement with a blind human, all claims; κ 0.48 on the real claims alone

Known defect

3/20 unanswerable questions got an answer instead of a refusal.

Python 3.12, FastAPI, Postgres + pgvector, Next.js, Claude.

Case study · Code

No. 02

PromptLang

Last frame of the bun run demo recording: semver-check rejects an incompatible change released as 1.1.0, passes it as 2.0.0, a consumer pinned to ^1.0.0 fails cleanly, then the 3 prompt tests pass on mock responses. — Last frame of the bun run demo recording. See the recording (GIF)

A typed language for LLM prompts, compiled to TypeScript or Python, published on npm.

96.13% mutation score (lexer, parser, semver; local run, outside CI)
44/45 AI linter recall; a linter that flags everything does as well

Known defect

12/40 AI linter flags judged correct in human review.

Strict TypeScript, Bun, npm.

Case study · Code

No. 03

Praxis

An LLM pipeline that turns a business question into an analytical briefing, with sourcing checks and adversarial review.

$3.29–$4.75 per completed live briefing

Known defect

6/6 live briefings flagged as insufficiently sourced.

TypeScript, Bun; Markdown, PDF and DOCX output.

Case study · Code

+ 02

Method

Pre-registered human review, published as is
Sample drawn with a fixed seed, committed before any annotation: 12/40 flags judged correct, published with its limit.
PR promptlang#39
Read the real outputs, not only the tests
My review of a Praxis briefing found estimates presented as facts. Validator fixed, review published.
PR praxis#47
Cold audit, every finding followed up in writing
An agent with no context audited PromptLang: 23 findings, each fixed, refuted or left open.
PR promptlang#32

All the rules and their incidents: how-i-build

+ 03

About

Software & AI Engineer in Lausanne, self-taught. I build AI products end to end and publish what they are worth, measures and defects.

/* Without the script the toggle cannot switch: show the detailed version and no inert control. */ [data-bio] .toggle { display: none; } [data-bio] [data-bio-version="long"][hidden] { display: block; }

A software and AI engineering role in French-speaking Switzerland? Write to me.

[email protected]

Write LinkedIn GitHub

I build AI products you can verify.

Projects

Verity

PromptLang

Praxis

Method

Pre-registered human review, published as is

Read the real outputs, not only the tests

Cold audit, every finding followed up in writing

About

A software and AI engineering role in French-speaking Switzerland? Write to me.