ReViva · AI for post-stroke rehabilitation

The problem

After a stroke, recovery depends on doing the right exercises, many times, correctly. Between sessions with a therapist, nobody is there to say whether a movement was right.

What we built

An app that watches rehabilitation exercises through a camera in real time and gives immediate feedback on every movement.

My part

I led a team from biology, AI, law and management, and built the machine-learning side: skeletal keypoints from the video stream, joint angles, features for pose recognition, the logic that scores each exercise, model testing and the integration into the app.

Result

1st at SYNLab, 3rd at EIT Health i-Days, 1st in Italy and 3rd internationally at the Sopra Steria Student Challenge, winner of the biomedical category at Startup Day, and the Longevity Award at Lifebility 2026.

A meeting agent that remembers

The problem

Teams keep making decisions their company has already ruled out: a legal blocker nobody remembered, an earlier agreement, a team that's already at capacity. Usually someone notices weeks later.

What we built

A B2B meeting agent that transcribes the conversation live and checks each decision against the company's memory, showing evidence-backed warnings while people are still talking.

How it works

Real-time transcription with Speechmatics, reasoning with Gemini 2.5 Flash, and retrieval over a graph of company knowledge, all running on Vultr infrastructure that survived some questionable load at 3 am.

Result

The prize for the Best Use of Gemini 2.5 Flash from Google and 2nd place for the Best Use of Vultr, at an event by lablab.ai during AI WEEK.

Product-query NLP service

The problem

People type grocery searches like “lactose free milk 2% for pancakes”. To show the right product, a shop has to understand the pieces (type, brand, volume, percent, attributes) despite typos, brand collisions and long-tail vocabulary, at production speed.

What we built

A production-style NLP service that splits a messy query into clean entities and links them to real SKUs in the catalogue.

How it works

A multilingual BERT fine-tuned for token classification; training data enriched with OpenFoodFacts and a parsed live catalogue; a lightweight post-processor that fixes tags and returns exact character spans for the API.

Result

13th of 61 teams by macro-F1, after late-night error hunts, ruthless evals and a team split across two countries.

Intent-aware retrieval for financial QA

The problem

Someone asking “what is a bond?” needs a different answer from someone asking “should I buy bonds now?”. Standard search ranks both in the same way.

What we built

A two-stage retrieval system over 57,000+ financial documents: dense retrieval with E5 and FAISS, then cross-encoder reranking.

My part

Classifying 648 queries into 6 intent types and building intent-aware reranking for definition, explanation and advice-seeking questions. We compared BM25, dense retrieval, hybrid search, HyDE query expansion and cross-encoder fine-tuning.

Result

+1.6% P@1 and +0.5% nDCG@10 over a strong dense retrieval + cross-encoder baseline.

Issa Pulire 2025 · IBM watsonx

The problem

Stadium cleaning turned out to be a surprisingly complex problem, and much of it is already driven by AI. The hard part was finding something truly new.

What we did

In six hours, with Maksim Okulov, Arseniy Kirykouski and Stefano Rizzuto, we dug into a niche we knew nothing about and pitched our idea on stage, with ISC S.r.l. sponsoring the team.

And then

A few weeks later, the IBM watsonx Agentic AI Hackathon: a completely new stack for our team, agents built on IBM watsonx Orchestrate.

What I took away

Even the most unexpected domains can use thoughtful, intelligent solutions, and a new stack is only scary until the first working demo.

AI content creator in n8n

The problem

Build an AI-powered content creator for professionals, with n8n. I had never used n8n before, and I signed up with one goal: learn a new skill in a day.

What I did

A day of learning, breaking and rebuilding, and a working solution shipped by the deadline.

Result

Top 11, and a place at the hackathon at Italian Tech Week 2026. Organised by Yellow Tech and Veliu.