Eleven works
At SerSan, the multi-agent system I built for the studio, a trading platform and its AI mentor, a workshop ERP, a WebGPU studio site, a real-estate platform and a brand film. Then Obla, my personal project, and three projects from outside SerSan.
-
01 / 11
SerSan · multi-agent system
SerSan Agent OS
Agents with roles, hand-offs under a brake and a Gate on every action that goes out; 2,513 tests passed.
Earned autonomy
each agent acts on its own only at the level it has earned
-
02 / 11
SerSan · live at spherenode.com
SphereNode
Vertical SaaS that replaced eight tools for the Italian retail-trading market.

-
03 / 11
SerSan · RAG mentor
Sphere AI
The bilingual mentor inside SphereNode, measured on a 161-question evaluation set.

-
04 / 11
SerSan · workshop ERP
Maintech OMA
Job orders on a six-phase state machine, ten trade portals.
9
roles, default-deny on every table
-
05 / 11
SerSan · Jun – Sep 2026
sersan.io
Seventeen scroll-locked scenes on one continuous WebGPU canvas.

-
06 / 11
SerSan · real-estate brand platform
Domus Tua
Mobile LCP 4.75 s → 1.74 s on real phones, no layout shift.

-
07 / 11
SerSan · identity and film
Brand film
Hexagonal mark, custom display typeface, colour system, and the brand film.

-
08 / 11
Personal project · multi-agent system
Obla
One chief of staff I talk to, and the agents it hands the work to, on my own machine.
Personal
my own multi-agent system, for me alone
-
09 / 11
ALS MCL Civitanova · freelance
Badante24h
Sub-second caregiver search; zero known auth vulnerabilities at launch.
Sub-second
geospatial caregiver search, PostGIS 3
-
10 / 11
DOIT S.r.l · internship
Voice AI Agent
Meetings and Dynamics 365 contacts, from natural-language speech.
Hands-free
appointments in Microsoft Teams, from audio notes
-
11 / 11
University of Camerino
Supply Chain
Ten roles, a team of three, on Spring Boot with JWT RBAC.

Measured before and after.
Two sets of figures with a before and an after: a 161-question bilingual evaluation set for Sphere AI, and page speed on real phones for Domus Tua.
Sphere AI · SerSan
Answer reliability, before and after
On a 161-question bilingual evaluation set. Citation correctness 100%, with zero locked-content leaks and zero fabricated citations.
Answer-policy accuracy
52%to68%
Over-answering on questions to withholdLower is better
56%to16%
Clarification precision
75%to94%
Clarification recall
32%to90%
Unsupported claimsLower is better
1.9%to0.6%
Too small to draw on this scale.
Domus Tua · real phones
Fast on the phone in your hand
The mobile delay was traced to the cookie banner and the intro rebuilt to run in CSS rather than JavaScript. The cinematic experience stayed, with no layout shift.
Mobile Largest Contentful PaintLower is better
4.75 sto1.74 s
Mobile blocking timeLower is better
572 msto274 ms