Pi School

Home
Italia
Roma, Lazio
Pi School

Indicazioni Sito Web

We help organisations turn AI into real business impact through applied research and practical training.

06/11/2026

𝐖𝐡𝐲 𝐛𝐮𝐢𝐥𝐝 𝐚𝐧 𝐀𝐈 𝐭𝐡𝐚𝐭 𝐬𝐩𝐞𝐚𝐤𝐬 𝐬𝐚𝐭𝐞𝐥𝐥𝐢𝐭𝐞 𝐜𝐨𝐦𝐦𝐮𝐧𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬?

On 26 November at 3:00 PM CET, our deep learning scientists Vijayasri Iyer and Riccardo Corrente will present "𝐵𝑢𝑖𝑙𝑑𝑖𝑛𝑔 𝑎 𝑝𝑟𝑖𝑣𝑎𝑡𝑒 𝐿𝐿𝑀 𝑓𝑜𝑟 𝑠𝑎𝑡𝑒𝑙𝑙𝑖𝑡𝑒 𝑐𝑜𝑚𝑚𝑢𝑛𝑖𝑐𝑎𝑡𝑖𝑜𝑛𝑠: 𝑡ℎ𝑒 𝑆𝑎𝑡𝑐𝑜𝑚𝐿𝐿𝑀 𝑐𝑎𝑠𝑒". They’ll walk through how to build and train the model and how to connect it to your own knowledge base.

The results are applicable to any organisation that possesses knowledge too specialised for general-purpose AI, or too sensitive to leave its own infrastructure.

👉 𝐑𝐞𝐠𝐢𝐬𝐭𝐞𝐫: https://pischool.link/SatLLM

Satellite communications engineering involves working with technical standards, link budgets, signal processing and calculations where precision matters. General-purpose language models, particularly smaller ones, can struggle with these tasks, producing incorrect formulas or hallucinating arbitrary values. This can have serious implications for the reliability of system design and analysis.

The European Space Agency's ARTES programme saw the collaboration of Pi School and RINA Consulting, leading to the creation of SatcomLLM. It is a private, satcom-specific assistant made up of two open models, a public corpus, human-written benchmarks, and a RAG system that answers from your own documents and cites them.

05/11/2026

𝗧𝗵𝗮𝗻𝗸𝘀 𝗳𝗼𝗿 𝗺𝗲𝗲𝘁𝗶𝗻𝗴 𝘂𝘀 𝗮𝘁 𝗘𝗠𝗡𝗟𝗣 𝟮𝟬𝟮𝟲
A week ago, our team presented on the System Demonstration Track at EMNLP 2026. The demo drew a steady stream of researchers and engineers, and over the following days Sébastien Bratières and Vijayasri Iyermet many of them to talk through how teams can measure coding agents on real refactoring work. The response confirmed there is a clear need for a reliable way to test these agents one variable at a time.
Thank you to everyone who stopped by.

👉 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫 𝐑𝐞𝐟𝐚𝐜𝐭𝐨𝐫𝐏𝐥𝐚𝐭𝐟𝐨𝐫𝐦: https://pischool.link/RefPlatfrm

03/11/2026

𝐓𝐡𝐞 𝐄𝐕𝐄 𝐭𝐞𝐚𝐦 𝐚𝐭 𝐭𝐡𝐞 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐈 𝐟𝐨𝐫 𝐄𝐚𝐫𝐭𝐡 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐭𝐢𝐨𝐧 𝐰𝐨𝐫𝐤𝐬𝐡𝐨𝐩

The Pi School team working on EVE was in Berlin for the 𝐴𝑔𝑒𝑛𝑡𝑖𝑐 𝐴𝐼 𝑓𝑜𝑟 𝐸𝑎𝑟𝑡ℎ 𝑂𝑏𝑠𝑒𝑟𝑣𝑎𝑡𝑖𝑜𝑛 𝑤𝑜𝑟𝑘𝑠ℎ𝑜𝑝, organised with BIFOLD - Berlin Institute for the Foundations of Learning and Data and ESA Φ-lab. During the workshop, they hosted the EVE EO MCP tool hackathon, where teams turned Earth observation services into tools EVE can use. Congratulations to the winners: OrbitCast by Cristopher Castro Traba, the Solenix EO toolset by Cesar Guzman and Matteo Zuracki, and the EO Spatial Split Planner by Stefano Francesco Pitton and Hugo Lee.

Àlex R. Atrio, Antonio Lopez, and Emanuele Micieli gave a live tour of EVE as an agentic Earth observation platform built on an open MCP tool registry. They explained how that registry is turning EVE from a single set of tools into an ecosystem, where the community can add tools and connect them to Earth observation data. With Nicolas Longépé of ESA Φ-lab, the team also led a session on testing Earth observation agents with existing benchmarks.

Thank you to BIFOLD, ESA Φ-lab and everyone who joined us. Follow EVE to stay updated on the project’s progress.

27/10/2026

𝗘𝗠𝗡𝗟𝗣 𝟮𝟬𝟮𝟲: 𝗖𝗼𝗺𝗲 𝗮𝗻𝗱 𝗳𝗶𝗻𝗱 𝘂𝘀 𝗶𝗻 𝗕𝘂𝗱𝗮𝗽𝗲𝘀𝘁.

On Sunday, Sébastien Bratières, our Director of AI, and Vijayasri Iyer, Senior Deep Learning Scientist, presented on the System Demonstration Track. They are around for the rest of the conference and happy to talk it through.

RefactorPlatform is an open-source evaluation harness for AI coding agents working on refactoring tasks. It grew out of our work with Tipico Group during School of AI session 17, where we built agents to restructure a codebase of over 2 million lines and realised there was no reliable way to measure how well they performed. The harness lets teams test one variable at a time, from the model to code search and agent architecture, with every run logged and exportable.

Across 100 multi-file refactoring tasks, giving the model structured code chunks rather than fixed-size ones increased accuracy by 25-30 percentage points. A single agent with good code search 86% outperformed a multi-agent setup 66% and improved retrieval accuracy without increasing the cost per successful task.

If you are bringing coding agents into your engineering team, look for Sébastien and Vijayasri in Budapest or explore the project: https://pischool.link/RefPlatfrm

23/10/2026

𝗪𝗵𝗮𝘁 𝗺𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗶𝗻 𝗔𝗜 𝗿𝗲𝗳𝗮𝗰𝘁𝗼𝗿𝗶𝗻𝗴?
If you are at EMNLP 2026, come and find out. Vijayasri Iyer, Senior Deep Learning Scientist, and Sébastien Bratières, Director of AI, are presenting our paper, 𝑅𝑒𝑓𝑎𝑐𝑡𝑜𝑟𝑃𝑙𝑎𝑡𝑓𝑜𝑟𝑚: 𝐴𝑛 𝑜𝑝𝑒𝑛-𝑠𝑜𝑢𝑟𝑐𝑒 ℎ𝑎𝑟𝑛𝑒𝑠𝑠 𝑓𝑜𝑟 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑑 𝑒𝑣𝑎𝑙𝑢𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑟𝑒𝑝𝑜𝑠𝑖𝑡𝑜𝑟𝑦-𝑠𝑐𝑎𝑙𝑒 𝑟𝑒𝑓𝑎𝑐𝑡𝑜𝑟𝑖𝑛𝑔 𝑎𝑔𝑒𝑛𝑡𝑠.

📍Find us at the main conference, or at our poster in Hall H, Session 2, on Sunday, 25 October, from 11:00 to 12:30.

One of our findings: more complex agent setups don’t always deliver better results. A single, well-designed agent capable of searching the codebase effectively outperformed a multi-agent setup, improving accuracy by 25 to 30 percentage points. This matters for engineering teams deploying AI coding agents. Before investing in a more elaborate architecture, it is worth checking whether the added complexity actually adds value.

Read the full article by following the link in the bio.

23/10/2026

20/10/2026

𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫 𝐑𝐞𝐟𝐚𝐜𝐭𝐨𝐫𝐏𝐥𝐚𝐭𝐟𝐨𝐫𝐦

𝐻𝑜𝑤 𝑑𝑜 𝑦𝑜𝑢 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 𝑤ℎ𝑒𝑡ℎ𝑒𝑟 𝑎𝑛 𝐴𝐼 𝑐𝑜𝑑𝑖𝑛𝑔 𝑎𝑔𝑒𝑛𝑡 𝑐𝑎𝑛 𝑟𝑒𝑠𝑡𝑟𝑢𝑐𝑡𝑢𝑟𝑒 𝑎 𝑐𝑜𝑑𝑒𝑏𝑎𝑠𝑒 𝑤𝑒𝑙𝑙?

Last year, while we were supporting our customer, Tipico, in solving a challenge their tech team faced, we found no controlled, reproducible way to measure the quality of results when an AI coding agent restructures a codebase.

So our team, Mann Acharya, Drish Mali, and Aziz Ben Amor, guided by Pi School Senior Deep Learning Scientist, Vijayasri Iyer, started building under the mentorship of Sébastien Bratières, our AI Director.

RefactorPlatform is an open-source evaluation harness for refactoring, one of the hardest tasks you can give a coding agent.

On 24 October, Vijayasri and Sébastien will present the paper at EMNLP 2026 in Budapest, one of the leading conferences in natural language processing research. We are proud to see the work accepted in a particularly competitive year.

Read more: https://pischool.link/RefPlatfrm

16/10/2026

𝐓𝐡𝐞 𝐧𝐞𝐱𝐭 𝐩𝐡𝐚𝐬𝐞 𝐨𝐟 𝐄𝐕𝐄: 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐭𝐡𝐚𝐭 𝐜𝐨𝐦𝐩𝐚𝐧𝐢𝐞𝐬 𝐜𝐚𝐧 𝐭𝐫𝐮𝐬𝐭.
EVE went public last week: the AI for Earth science that Pi School built under the Foresight element of ESA’s FutureEO programme, with a €2 million investment from ESA.
𝐍𝐨𝐰 𝐰𝐞’𝐫𝐞 𝐭𝐚𝐤𝐢𝐧𝐠 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐬𝐭𝐞𝐩, 𝐟𝐫𝐨𝐦 𝐚𝐧 𝐀𝐈 𝐭𝐡𝐚𝐭 𝐚𝐧𝐬𝐰𝐞𝐫𝐬 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬 𝐭𝐨 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐭𝐡𝐚𝐭 𝐠𝐞𝐭 𝐰𝐨𝐫𝐤 𝐝𝐨𝐧𝐞.
From 19 to 21 October, our team will be in Berlin at the Agentic AI for Earth Observation workshop, with a live demo of EVE, a talk on the tool registry, and two hands-on sessions.

On 22 October, with BIFOLD - Berlin Institute for the Foundations of Learning and Data and ESA Φ-lab, we’re running a one-day hackathon on EVE’s open tool registry. In a single day, teams of scientists and engineers will connect satellite data services to AI agents and turn them into working results: a flood map, a forest regrowth report, a climate briefing. Every step is traced and every source cited.

That’s the approach we bring to every project. Agents built on open standards avoid vendor lock-in, allow tools to be reused across teams, and enable every output to be checked. It’s what makes AI usable in organisations where accuracy matters.

𝐈𝐟 𝐲𝐨𝐮’𝐫𝐞 𝐢𝐧 𝐁𝐞𝐫𝐥𝐢𝐧, 𝐥𝐞𝐭’𝐬 𝐦𝐞𝐞𝐭. Try EVE: https://eve.philab.esa.int

11/10/2026

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟏𝟎𝟎 𝐢𝐬 𝐡𝐞𝐫𝐞!

It’s Friday! Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Senior Deep Learning Scientist, Àlex R. Atrio.

This week’s highlights:

🔎 𝐉𝐞𝐯, 𝐋𝐚𝐲𝐚 & 𝐎𝐩𝐞𝐧-𝐉𝐞𝐯: 𝐀𝐈 𝐓𝐡𝐚𝐭 𝐃𝐞𝐜𝐢𝐝𝐞𝐬 𝐈𝐧𝐬𝐭𝐞𝐚𝐝 𝐨𝐟 𝐖𝐫𝐢𝐭𝐢𝐧𝐠: TypeSafe AI's Jev is a "System One" model that doesn't generate text. It takes a state and a set of typed questions, and returns calibrated decisions in milliseconds at a fraction of LLM cost. Within days, two open-weight alternatives followed: Convai Innovations' Laya and the community-built Open-Jev. Think of them as smart if-statements for AI agents.
🌐 https://pischool.link/8ab

🧠 𝐖𝐑𝐎𝐏: 𝐓𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝐖𝐨𝐫𝐥𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 𝐎𝐛𝐣𝐞𝐜𝐭 𝐏𝐞𝐫𝐦𝐚𝐧𝐞𝐧𝐜𝐞: Video world models often make hidden objects vanish, morph or teleport. Yilun Du, Lvmin Zhang (the creator of ControlNet) and Haotian Zhang tackle this with two releases:
WROP: a benchmark of 150 cognitive tasks.
PWM-WROP: a 16B physical world model that cuts spatial deformation errors from 64.3% to 12.4%.
Weights, code and data are all open. This was the top paper on Hugging Face this week.
🌐 https://pischool.link/d55ea1

🤖 𝐀𝐠𝐞𝐧𝐭𝐖𝐨𝐫𝐥𝐝: 𝐂𝐚𝐧 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭𝐬 𝐀𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐖𝐨𝐫𝐤 𝐚𝐬 𝐚 𝐓𝐞𝐚𝐦?: Most multi-agent benchmarks test competition or short exchanges. AgentWorld tests real teamwork instead: 100 tasks in a game-style sandbox where teams of 3–20 agents with different roles have to communicate, plan and share resources over 50+ rounds. The best models succeed only 52% of the time, failing through communication breakdowns, role confusion and plans that fall apart. The authors also propose a new metric that measures how much of a team's effort actually contributed to the result.
🌐 https://pischool.link/101969

Was this helpful? Let us know by liking and sharing!

09/10/2026

𝐄𝐕𝐄 𝐢𝐬 𝐧𝐨𝐰 𝐨𝐩𝐞𝐧 𝐭𝐨 𝐞𝐯𝐞𝐫𝐲𝐨𝐧𝐞.
The Earth Virtual Expert answers complex Earth science questions in plain language, with every answer traceable to scientific sources. Anyone can create a free account and start today.

EVE has been developed under the Foresight element of ESA’s FutureEO programme, with a €2 million investment. Pi School led the design, development and integration, working with ESA, Mistral, Wiley and Imperative Space.
"𝐼 ℎ𝑎𝑣𝑒 𝑎𝑙𝑤𝑎𝑦𝑠 𝑏𝑒𝑙𝑖𝑒𝑣𝑒𝑑 𝑡ℎ𝑎𝑡 𝑡ℎ𝑒 𝑚𝑜𝑠𝑡 𝑚𝑒𝑎𝑛𝑖𝑛𝑔𝑓𝑢𝑙 𝐴𝐼 𝑖𝑠 𝑡ℎ𝑒 𝑘𝑖𝑛𝑑 𝑡ℎ𝑎𝑡 𝑠𝑒𝑟𝑣𝑒𝑠 𝑠𝑜𝑐𝑖𝑒𝑡𝑦." Isabelle Andrieu, CEO of Pi School.

Try EVE: https://eve.philab.esa.int

ESA Earth Observation

Indirizzo

Via Indonesia 23
Rome
00144

Sito Web

http://www.picampus-school.com/

Notifiche

Lascia la tua email e sarai tra i primi a sapere quando Pi School pubblica novità e promozioni. Il tuo indirizzo email non verrà utilizzato per nessun altro scopo e potrai annullare l'iscrizione in qualsiasi momento.