Principal Data Scientist at Endava. Kaggle 2xGM. GDE. Engineer. Christian. Father. Passionate about Literature, History and Mountains. Opinions my own.
Joined July 2013
- Tweets 3,160
- Following 328
- Followers 2,062
- Likes 2,391
343 Photos and videos
May 27
I just published: Introduction to Agentic AI with Google ADK 🚀 on @towards_AI (@Medium): lnkd.in/d2ctaemU
Learn how to build AI agents with Gemini that can:
• use tools
• orchestrate workflows
• collaborate with agents
• deploy on GCP
#AI #AgenticAI #ADK #A2A
93
May 25
I just published on @Medium the article "From Vertex AI to Gemini Enterprise Agent Platform": medium.com/p/from-vertex-ai-…
#AgentPlatform #AgenticAI #AgentEngine #ADK #A2A #GCP #GDE
62
May 9
I just published Running MedGemma on Ollama: Multimodal Medical AI in Action on @towards_AI @Medium publication: pub.towardsai.net/running-me…
#MedGemma #Gemma #Medical #AI #Healthcare #MedicalImage #GCP #GoogleCloud #TowardsAI
1
5
213
May 5
I just published "What reborn Kaggle teaches us about the future of AI" medium.com/p/what-reborn-kag…
#Kaggle #AI #AISystems #GameArena #Benchmarks #FreeTier
1
68
Apr 30
🚀 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠 𝐆𝐞𝐦𝐦𝐚 𝟑 𝟐𝟕𝟎𝐌: 𝐌𝐞𝐝𝐢𝐜𝐚𝐥 𝐐&𝐀 𝐰𝐢𝐭𝐡 𝐊𝐞𝐫𝐚𝐬 & 𝐉𝐀𝐗
🔗Medium: lnkd.in/dJ5K4mm8
🔗 Medium Friendly link: lnkd.in/dZSKbt_F
🔗 GitHub: lnkd.in/da6SQseW
#Gemma3 #Keras #JAX #LLM #M #AI #GCP #TPUSprint #GPU #LoRA
1
2
149
Apr 20
I just published Getting Started with Gemma 4: A Practical Guide to Multimodal, Multilingual AI medium.com/p/getting-started…
2
139
Apr 19
I just published Run Gemma 4:E4B Locally with Ollama — No Cloud, No Limits, Real Results medium.com/p/run-gemma-4-e4b…
Friendly link: medium.com/@gabi.preda/run-g…
1
180
Apr 18
Why should Romanian politics be a choice between skilled flamboyant thieves, boring yet kind of competent bureaucracy, and delusional pro-Russian traitors?
1
69
Apr 14
I ran Gemma 4 E2B on CPU only.
No GPU. No tricks. Just a Kaggle notebook (with CPU). No quantization.
🔗 Full @Medium @towards_AI write-up: lnkd.in/dSui726p
🔗 Friendly link on @towards_AI: lnkd.in/dh4eEN3N
🔗 Kaggle Notebook: lnkd.in/dmVPUTEw
#Gemma4 #CPU
5
302
Apr 11
I just published Running Gemma 4:E2B on Android: A Minimal Kotlin App medium.com/p/running-gemma-4…
1
111
Apr 7
I just published From OOM Errors to Working Model: Fine-Tuning Gemma 4 E2B Step-by-Step using Unsloth medium.com/p/from-oom-errors…
2
104
Apr 7
I just published Run Gemma 4:E2B Locally with Ollama: No Cloud, No Limits medium.com/p/run-gemma-4-e2b…
1
83
Apr 7
I just published Getting Started with Gemma 4: A Practical Guide to Multimodal, Multilingual AI medium.com/p/getting-started…
3
70
Apr 2
I just published Getting Started with Gemma 4: A Practical Guide to Multimodal, Multilingual AI medium.com/p/getting-started…
1
67
Gabriel Preda retweeted
Mar 22
see this ➡️ “Developing Kaggle Notebooks, toward becoming a @Kaggle Notebooks Grandmaster” at: amzn.to/48LAjKA by @PredaGabi v/ @PacktDataML
——————
#DataScience #MachineLearning #ML #DataScientist #Python
2
6
814
Is a high accuracy score on static benchmarks enough to trust AI in the real world?
Static accuracy is no longer enough. @PredaGabi explores how Kaggle Benchmarks are bridging the gap between the leaderboard and the real world. 🧵
1
2
25
5,690
Mar 2
Romanian History Bias&Integrity Benchmark (RHBIB) on #kaggle kaggle.com/benchmarks/gpreda…
2
82
🌟 Kaggle Community Spotlight!
Can LLMs handle the nuance of historical facts or do they just repeat popular myths? 🏛️
Kaggle Grandmaster @PredaGabi uses the new Kaggle Community Benchmarks to put leading models to the test on complex, disputed topics.
Learn more: medium.com/@gabi.preda/how-g…
ALT A screenshot of a leaderboard titled "Does LLMs know history?" comparing the performance of four AI models across ten historical tasks. The models—Gemini 3 Pro Preview (0.90), Claude Sonnet 4 (0.80), DeepSeek-R1 (0.80), and Gemini 2.5 Flash (0.80)—are displayed in a grid showing "PASS" and "FAIL" results. The tasks cover topics such as the Byzantine Empire, the Roman Empire, and Joan of Arc. Notably, Gemini 3 Pro Preview is the only model to pass the "last_egypt_pharaoh" task, while all four models failed the "vestal_virgin_jurisdiction" challenge. The interface is clean, featuring a white background with green and red status indicators for each task.
4
32
4,729