Daniel Svonava

Superlinked

superlinked.com makes it easier to turn data into vector embeddings

Co-Founder & CEO

creable.io

creable.com is the world's biggest influencer database (300M creators)

Advisory Board Member

Tater Games

tatergames.com is a data-driven game studio

Board Member

Primeflow

primeflow.com built tools for super-connectors and their network-driven business

Founder & CTO

Google

2016 - 2018
Tech-lead for YouTube Reservation Ads forecasting (Can a reservation be booked or not?) and pricing (How much should we charge for it?).

2015
Tech-lead for YouTube Video Ads traffic prediction and cross-Google Ads reach prediction, enabling ad buying in Adwords and DBM.
Focus on making Google Ads useful for Brand advertisers.

2013 - 2014
Within YouTube monetization I'm constantly switching between being an analyst who organizes and makes sense of data, builds prototype visualizations and spreads around the insights to help decision making and product design and being an engineer who uses these insights to build useful production systems.

Senior Software Engineer

47pixels

We were a software startup in the Story-telling space. Major products:
47posters - B2B SaaS for Print Service Providers launched in 7 countries
memoarc - B2C storytelling photogallery
dofsome - B2C minimalistic, high-quality photo hosting.
Reach out to me for a complete portfolio.

Co-Founder

Exploratory large-scale information mining from unstructured data. Patent granted.

Software Engineering Intern

IBM Research

Joined the Information analytics group within the Mathematics and Computer Science department again to continue my work on matrix reordering and co-clustering remotely via IBM SK.

Research Internship

Joined the Information analytics group within the Mathematics and Computer Science department. Co-authored a research paper involving graph theory, clustering and visualization. Worked on a web-based marketing intelligence tool employing data-mining techniques and back-end infrastructure optimization.

"Great Minds" Research Internship

Slovak University of Technology

The course is called Development of effective algorithms and programs. I prepared problem statements and solutions for assignments and competitions. I also set problems for ProFIIT, national programming competition for High school students.

Teaching assistant

Vector Compute @ Superlinked | xYouTube

➡️ Building a compute framework for turning data into vector embeddings.
🤗 Connect your data infrastructure to your vector database with Superlinked.

Join our product preview:
📋 Try it out in a Python notebook with no dependencies.
↗️ Design the perfect vectors for your retrieval use-case.
🔧 Mix and match streaming, batch, structured and unstructured data.
🚀 Launch to production with your favorite Vector Database.

➡️ Let's sort out information retrieval across your organization!
.
.
.
.
Previous:

Daniel is a Senior Software Engineer at Google, where he tech-leads, designs and builds forecasting & pricing systems for ads worth billions of dollars. These are distributed systems that run on thousands of machines, process terabytes of data per query and combine Machine Learning modeling with game-like simulation of tens of thousands ad campaigns - all happening in real-time to enable thousands of concurrent users of buying tools like AdWords and DBM.

Daniel wields a rare combination of deep technical skills, a user-centric drive and an ability to align multiple teams towards a common goal - launching and landing high quality cutting edge products.

Daniel co-founded a tech startup that brought machine vision to the consumer photo space 10 years ago, he was an invited speaker at an ML dev conference in Zurich and speaks at startup events regularly. He holds multiple patents with Google, is published in a peer-reviewed ML journal from his time at IBM Research and is a Google Code Jam on-site finalist.

On weekends he flies, climbs, skis and bikes in the Alps, see: http://instagram.com/mountain.supo

Toptal

Machine Learning engineer / Data Scientist

Braintrust

Machine Learning Engineer | Founding Talent Member

EasyHealth

Machine Learning Engineer

Cogsy

Time series ML engineer

Speedy Mentors

Mentor

Fifteenkay

Machine Learning Consultant

Speak

Data Engineer

Lola Market

- Developed, deployed, and maintained a Machine Learning model to improve the efficiency of the shoppers' fleet.
- Bootstrapped the first data-warehouse and reporting layer in the company (Amazon Redshift + Amazon DMS + Tableau).

Iridium Advisors

- Statistical analysis of financial markets using explainable Machine Learning techniques.
- Benchmarked a range of ML solutions to enhance model accuracy and explainability.
- Wrote a Python package to ensure the in-house reproducibility of each step of the analysis (data processing, data validation, data visualization, model construction, model validation, and model explanation).

cyngn

- Developed and maintained the analytics dashboards for Cyngn's engineering team and Cyngn's clients.
- Developed components of the internal ETL tool in Python and SQL.
- Assisted the back-end engineer to integrate both frontend and backend of the analytics stack inside Amazon Redshift.

Data scientist

Author of the Hands-on Reinforcement Learning course 🚀

Data scientist and ML engineer with over 8 years of experience. I have a passion for building ML-based solutions, from development to deployment. I love transforming an idea into a model, and a model into an API/product.

I have worked on different problems: financial derivative pricing, digital marketing analytics, deep learning for art generation, or demand prediction for on-line shopping. I have a background in pure mathematics and strong coding skills in Python.

Pau Labarta Bajo

Tip for AI engineers 💡

If you want to reduce your cloud bills by 80% you can replace your god-like model with a router + SLM where

- 80% of the tasks are handled by a small language model
- 20% of the tasks are sent to your large/powerful/expensive model.

You are welcome

Building and teaching AI that works ​> Maths Olympian​> Father of 1.. sorry 2 kids

The future of agentic AI is hybrid and smaller models are the ones calling the shots ↓
  
  
A couple of weeks ago NVIDIA released Orchestrator-8B

Orchestrator-8B is a state-of-the-art 8B parameter orchestration model designed to solve complex, multi-turn agentic tasks by coordinating a diverse set of expert models and tools.

In other words, this model helps you optimally combine:

> Basic tools, like internet search or a sandboxed code interpreter
> Specialized LLMs, like Qwen2.5-Math-7B
> Frontier LLMs like GPT-5, Claude Opus 4.1 and Gemini

to solve hard problems, while optimizing for speed, accuracy, AND cost.


𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀? 🤯

On the world's hardest benchmarks:
> 37.1% on Humanity's Last Exam (vs GPT-5's 23.4%)
> 76.3% on FRAMES (vs GPT-5's 66.3%)
> 80.2% on τ²-Bench (beating all SOTA models)

At a fraction of the cost:
> $9.2 average cost (vs $30.2 for GPT-5, $52.5 for Claude Opus 4.1)
> 8.2 latency score (vs 19.8 for GPT-5, 25.6 for Claude Opus 4.1)


𝗪𝗵𝘆 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝘄𝗼𝗿𝗸?

The counterintuitive magic: small models excel at orchestration precisely because they're unburdened by excessive knowledge and trained to capture the essence of problem-solving.

They act as intelligent coordinators by calling the right expert at the right time, rather than trying to do everything themselves.

This is trained using ToolOrchestra, a method combining:

-> Synthetic data generation
-> Multi-objective reinforcement learning
-> Cost-aware optimization


𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗶𝗰𝘁𝘂𝗿𝗲

This represents a shift from monolithic LLMs to compound AI systems, where specialized models work together under smart coordination.
The result? Better performance, lower costs, and the scalability needed for real-world agentic AI.

----
Follow me for more high-signal (and less BS) on your feed.

Small Language Models aren't just for edge devices. They're silently cutting cloud costs by 80%+ for production workloads. Here's the math 🧵👇
  
    
Most people think: Small models = edge deployment only.

But here's what they miss: Cloud inference costs are DOMINATED by compute time, not storage.

A 1B param model can be 10-20x cheaper to run than a 70B model on the same hardware.


𝗘𝘅𝗮𝗺𝗽𝗹𝗲

You need to process 10M API calls/day for classification or simple reasoning tasks.

70B model: ~$X per 1K tokens
3B model: ~$X/10 per 1K tokens

That's potentially $XXXk saved per month. 💰

The secret sauce? Task-specific optimization.
You don't need GPT-5 level intelligence for 80% of the problems:

> Content moderation
> Simple Q&A
> Data extraction
> Classification tasks
> Routing queries

A fine-tuned 3B model often outperforms a generic 70B here.


𝗣𝗿𝗼 𝘁𝗶𝗽: 𝗨𝘀𝗲 𝗮 𝗿𝗼𝘂𝘁𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻

→ Small model handles 80% of simple queries
→ Large model only for complex reasoning
This gives you the best of both worlds while slashing costs.


TL;DR: Stop thinking of small models as "just for mobile."

They're your secret weapon for profitable AI products in the cloud.

Start with the smallest model that can do the job. Scale intelligence only when needed.

You are welcome.

Are you someone who checks loss curves before checking your phone in the morning?
  
  
Then Liquid AI Labs is your place.

We're hiring ↓

Open article: Liquid AI Jobs by jobs.ashbyhq.com

If you spent the week on Saturn and just got back to Earth, this is the technical report by Liquid AI that explains how efficient AI models are built.

You are welcome.

Open article: LFM2 Technical Report by arxiv.org

If you spent the week on Mars and just got back to Earth, this is the blog post you should start reading by the Hugging Face team ↓↓↓

Open article: Evaluation Guidebook - a Hugging Face Space by OpenEvals by huggingface.co

Evaluation Guidebook - a Hugging Face Space by OpenEvals

I fine-tuned a Visual LM for a complex image classification task.
Here's what I learned ↓
  
  
𝗪𝗵𝗮𝘁 𝘄𝗮𝘀 𝘁𝗵𝗲 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲?
Two weeks ago I fine-tuned LFM2-VL-450M by Liquid AI to perfectly identify images of cats vs dogs. That was my first victory, but it felt too easy.

Hence, on Friday last week I decided to solve a problem closer to the real-world.

I chose the Stanford Cars dataset, which contains roughly 7,000 images of cars with their exact car maker, model and year.

Let me show you what I did


𝗦𝘁𝗲𝗽 𝟭 -> 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝘁𝗵𝗲 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗮𝗻𝗱 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀
To simplify things a bit, I stripped only the car maker, so the problem to solve was a 49-class classification problem from images.

I split the data into disjoint training and evaluation datasets to perform fine-tuning without data contamination.


𝗦𝘁𝗲𝗽 𝟮 -> 𝗘𝘀𝘁𝗮𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗮 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆
I tested the accuracy of 3 Visual Language models for this task, without any fine-tuning. 

> LFM2-VL-450M -> 60%
> LFM2-VL-1.6B  -> 72%
> LFM2-VL-3B -> 80%

The latest 3B parameter model released by Liquid AI is a small beast, that already sets a high-bar on this difficult task.


𝗦𝘁𝗲𝗽 𝟯 -> 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝘁𝗼 𝘀𝗾𝘂𝗲𝗲𝘇𝗲 𝗮𝗹𝗹 𝘁𝗵𝗲 𝗷𝘂𝗶𝗰𝗲
Using LoRA for efficient fine-tuning, I trained across 500 steps and tracked intermediate checkpoints.

Observations:

> The train loss curve for each model saturates at very different levels (see image attached to this post). Both the 450M and 1.6B models struggle to fit on the training data, while the 3B seems to manage it well. Based on this, I picked the 3B for further analysis.

> The eval loss curve of the 3B model is decreasing up to 500 steps, meaning not overfitting on the training data. Good.

> I saved checkpoints every 100 steps so I could run evaluations on intermediate checkpoints to better understand the model’s performance at different stages of the training process.

𝗥𝗲𝘀𝘂𝗹𝘁𝘀
- After 100 steps -> 𝟱𝟮% accuracy, worse than baseline.
This is because the LoRA parameters are still to noisy and damage the model performance.

- After 200 steps -> 𝟴𝟰% accuracy, surpassing the original.
The model is already "in the money" and performs better than the baseline.

- At 500 steps -> 𝟴𝟲% accuracy, approaching 90%.

I did not have time to fine-tune further, but my intuition is that as I cover a whole epoch on the training dataset, I will start to overfit, and the evaluation accuracy will not get close to what I want (aka 99.9% accuracy)

𝗦𝗼, 𝘄𝗵𝗮𝘁'𝘀 𝗻𝗲𝘅𝘁?
I will continue the training process, to confirm my intuition and (if I have time) I will try to do some data engineering to get more samples per class and improve the model’s accuracy.

Let’s keep on VLM-fine tuning, and follow me to get more hands-on content like this.

All code is open-source on GitHub ⭐ You can find the link in the comments ⬇️⬇️⬇️

Open Science is what this planet needs 🌎🌍🌏
Thank you Liquid AI for doing this ↓

Today we introduce Liquid Labs, our advanced research unit, with the goal of understanding and building efficient and adaptive intelligence systems.

Liquid Labs consolidates our existing research efforts at Liquid across architecture of foundation models, multimodality, training, data, and inference. 

The lab also will be home to new frontier research work across the broad range of foundation model build-up stack.

Read the full announcement: https://lnkd.in/ezChkkGD

We are hiring: https://lnkd.in/eYKRC5sH

Also find us at NeurIPS 2025 exhibition hall! 🚀

The most underrated paper of the year
arxiv.org/pdf/2511.21689

Open article: ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration by arxiv.org

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Mathematical certainty, not institutional trust.

I work at the intersection of AI and cryptography.

Language barriers affect almost everything in international business.. but they are about to disappear. 

In 2026 there will be zero excuses why every single doc in your org is not instantly accessible by everyone in your org (regardless of language)!

Here's one example from my past. A problem killing many Japan startups quietly:  In Japan, only ~15% of people read English proficiently. 
Globally? About 20%.

Yet 95% of cutting-edge AI research, technical documentation, and developer resources are published in English first.

This creates an invisible ceiling. 

Japanese startups can't access the knowledge they need if they don't also have English proficient talent.. Found the best dev? Probably not going to have the best english. Getting the best dev who is also fluent!? Super hard. 

I know I have run startups in Japan. It's a vicious cycle. 

But what if language barriers just... disappeared?

I've been experimenting with Liquid AI 's LFM2-ColBERT-350M — a late interaction retriever that does something super cool:

Upload docs in English → Query in Japanese (or any of 8+ languages) → Get accurate results. 

Any knowledge base can instantly be accessible to your JP team — at basically no cost!

Your Japanese engineers can now search English technical docs in their native language. Your French team can query your English knowledge base. 

The research gap closes. But there's a trust problem. Does every company need to in-house index all potentially valuable docs into their own vector DB? Or can we trust one another and share?

RAG systems are only as good as the data you feed them. How do you prove your knowledge base isn't poisoned by a competitor, outdated info, or manipulated content?

This is where zkML comes in.

Zero-knowledge machine learning lets you cryptographically prove: 
✓ Only verified sources entered the database 
✓ No poisoned data contaminated your embeddings
✓ The retrieval model ran correctly without tampering 
✓ All without revealing the actual data

Imagine: A Japanese developer queries your system in Japanese, retrieves English documentation, and can cryptographically verify that every piece of information came from legitimate, unaltered sources.

The implications:
For regulated industries (healthcare, finance, legal) → Provable compliance across languages.

For open-source communities → Verifiable knowledge bases anyone can trust. 

For global teams → One source of truth, accessible in any language, with mathematical guarantees.

We're not just breaking down language barriers. We're building trustless, multilingual knowledge systems that scale globally.

The fabled Tower of Babel scattered us with confusion...
Modern tech is bringing us back together w/ cryptographic receipts.

Check out JOLT Atlas! This plus LFM2-ColBERT-350M is used in our PoC to do  RAG that can be used across orgs without trust!

https://lnkd.in/gticfE8e

#zkML #AI #RAG #MultilingualAI #TechInnovation #StartupChallenges #ZeroKnowledge #LiquidAI

Open article: GitHub - ICME-Lab/jolt-atlas: Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡ by github.com

Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡ - ICME-Lab/jolt-atlas

GitHub - ICME-Lab/jolt-atlas: Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡

taplio

The Battle

Daniel Svonava

Pau Labarta Bajo

Followers

Posts

Engagements

Likes

Performance ScorePau Labarta Bajo's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Daniel Svonava

Pau Labarta Bajo

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Pau Labarta Bajo's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePau Labarta Bajo's LinkedIn Performance