Daniel Ionut Vasile
Greater London, England, United Kingdom
188 followers
178 connections
188 followers
178 connections
Activity
188 followers
-
Daniel Ionut Vasile reacted on thisGracias Antonio Larronda y a El País Uruguay por la nota! Desde que arrancamos con Ignacio De Haedo en enero 2025, Mirando pasó de una idea de usar IA con datos inmobiliarios a algo que cada vez inmobiliarias usan todos los días para generar una mejor experiencia para sus clientes. En el medio, muchas iteraciones, el apoyo de ANII, nuestro primer hire Juan Francisco De Haedo Falco, y excelentes clientes que nos enseñan mucho. Estamos acelerando y con nuevos productos que no llegamos a hablar en la nota. Si te interesa sumarte a un equipo chico construyendo productos para LATAM escribime (tech o no tech) y charlamos.Daniel Ionut Vasile reacted on this🤯 ¡Los uruguayos que renunciaron a Meta en Londres, emprendieron con IA para inmobiliarias y ya desembarcaron en Argentina! 💡 Ignacio De Haedo y Javier Rey trabajaban en Meta, en Londres. Pero en 2024 decidieron que era hora de hacer algo propio. 🤖 Renunciaron, comenzaron a probar hasta que llegaron a un producto, una plataforma que aplica IA para búsquedas inmobiliarias. 🏘️ Tras algunas pruebas a inicios del año pasado, en agosto de 2025 lanzaron Mirando, una herramienta dirigida para operadores inmobiliarios que automatiza la búsqueda de propiedades y optimiza la gestión. 🚀 El rápido exito, junto a que obtuvieron financiamiento de ANII, hicieron que en febrero de este año desembarcaran en Argentina. ¿Más detalles? 📰 Mi nota en #ElEmpresario de El País Uruguay 👉https://lnkd.in/dKHpRzJt
-
Daniel Ionut Vasile liked thisDaniel Ionut Vasile liked thisHiring in Warsaw for a Staff Backend/AI Engineer, CS Automation at Bolt. We are developing the agentic AI infrastructure that supports customer service for over 200 million users in more than 50 countries. This role involves multi-agent orchestration, LLM evaluation pipelines, guardrails, and observability at a production scale. You will be responsible for the technical vision and architecture across multiple teams, working alongside a talented group of engineers, data scientists, analysts, product managers, and more. We are looking for data-driven candidates with hands-on experience in Machine Learning and/or agentic AI patterns and strong skills in distributed systems. This is not a wrapper-around-an-API role, but a full AI platform one. If you are eager to make large AI systems function effectively in production, please reach out or apply at the link below: https://lnkd.in/dtzHKWQF Note: while the opening is in Warsaw, if you fit the profile and are in one of our engineering hubs (Tallinn, Bucharest, Berlin, Tartu), reach out to me regardless and let's discuss about opportunities at Bolt :) #hiring #staffengineer #agenticai #boltStaff Backend Engineer, AI Applications | Bolt CareersStaff Backend Engineer, AI Applications | Bolt Careers
-
Daniel Ionut Vasile reacted on thisDaniel Ionut Vasile reacted on thisI’m excited to share that I’ve been promoted to Senior Engineer (IC5) at Meta. Over the past 2.5 years, I’ve grown from an Intern to a Senior Production Engineer overseeing Site Reliability for Ads ML Systems. Grateful to the mentors and teammates who’ve guided me along the way. Looking forward to driving greater impact!
-
Daniel Ionut Vasile reacted on thisDaniel Ionut Vasile reacted on thisObserving human behaviour, 2025 edition. Patterns repeat.
-
Daniel Ionut Vasile liked thisDaniel Ionut Vasile liked thisA/B-tests are everywhere, and with good reason. Whether you want to test a new product feature, a UX change, or a machine learning model, randomised controlled experimental designs bring a ton of advantages. Once all the data is collected, the statistical analysis is often reasonably “out of the box”: estimate a confidence interval on the Average Treatment Effect (ATE) for metric Y, and if the interval is strictly positive, exclaim statistical significance and pop the champagne. The Central Limit Theorem (CLT) is responsible for ease-of-use here: because the sampling distribution of the ATE will approach normality as we get more samples, we can get confidence intervals by simply computing the mean and standard error to fit a Gaussian, and we’re done… Or are we? The CLT guarantees that this works when the sample size is “large enough”. But what exactly does this mean? ChatGPT and DeepSeek guess 30, with a caution that skewed metrics might need 50, 100, or more. Spoiler alert: it’s more. Because behavioural data often follows power-law dynamics, we should expect distributions to be skewed, and we should be skeptical about the methods we use! At Aampe, we’ve observed that some event metrics with thousands of occurrences across millions of users still exhibit non-normally distributed ATEs. The result is that a simple 95% or 99% confidence interval loses its meaning, as we lack any and all guarantees that it is actually likely to include the true ATE. If we can pinpoint such cases early on, that allows us to consider other metrics, methods, or sample sizes. And that leads to more robust decision-making, which definitely benefits businesses. It turns out that we can test these –often implicit– assumptions by repeatedly resampling A/A-tests and performing a Kolmogorov-Smirnov test on the distribution of the measured ATEs, or their resulting p-values. This is effective, efficient, and helps uncover problematic experimental designs before disaster occurs! A technical write-up on our approach is now on arXiv, URL in the comments.
-
Daniel Ionut Vasile reacted on thisI could not have asked for a better way to end my PhD, or enjoy my first physical conference in two years!Daniel Ionut Vasile reacted on thisProud of my student Olivier Jeunen for winning the student best paper award at the ACM conference on Recommender Systems! #Recsys2021 #uantwerpen
-
Daniel Ionut Vasile liked thisNot to mention the prediction horizons of up to 1h, making it an MLOps nightmare to use such models in real time :)Daniel Ionut Vasile liked thisDeep Learning SotA papers are like 15 model stacked ensembles on Kaggle - kind of impressive, not always practical. Here's last week's example (https://lnkd.in/g6iDiutK) from DeepMind et. al, where they model traffic using Graph Neural Networks. The authors' own conclusion: "Our model (GN) significantly outperforms all reported baselines for each region and horizon" What's actually in the paper: * GN outperforms a model without any graph structure by roughly <1% (sic) of RMSE * Road segment embeddings are similarly low impact, improving RMSE 0-2%, road supersegment embeddings are even detrimental at times * Only trivial benchmarks considered - purely historical traffic, purely real time traffic To be honest, this makes me wonder if a silly Random Forest/linear model using both historical and real time features might achieve similar performance to the elegant graph model, with perhaps 10% of the development/deployment effort :) Investing writing time into realistic benchmarks is the best way to signal the actual superior performance of a novel method. #publicationbias #deeplearning #ai
-
Daniel Ionut Vasile reacted on thisDaniel Ionut Vasile reacted on thisA few months ago, I saw Lihong Li tweet about an “Early-Career Scientist” programme at Amazon Science. Seemed like a long shot, but I decided to try my luck. Two months from now, I’ll be joining Amazon as a Post-Doctoral Scientist in Edinburgh! Can't wait for the opportunity to work with Ben Allison, Lihong Li and the team at Amazon for the following two years. Exciting times!
Experience
-
-
-
-
-
-
-
-
-
Education
-
-
-
-
Licenses & Certifications
Projects
-
Personalised learning-to-rank for e-commerce search
-
See projectProposed a personalised ranking solution for e-commerce search as part of my Bachelor’s project. The developed algorithm is based on an ensemble of rankers modelled by an adaptation of RankNet. Starting from a set of raw search-engine logs, the project covered feature extraction, designing and training the ranking ensemble, building a validation framework for it, and ultimately evaluating it by submitting it on the test leader board of the ACM CIKM Cup 2016 competition, on which it occupied the…
Proposed a personalised ranking solution for e-commerce search as part of my Bachelor’s project. The developed algorithm is based on an ensemble of rankers modelled by an adaptation of RankNet. Starting from a set of raw search-engine logs, the project covered feature extraction, designing and training the ranking ensemble, building a validation framework for it, and ultimately evaluating it by submitting it on the test leader board of the ACM CIKM Cup 2016 competition, on which it occupied the 4th position.
-
Oxercise: Wellbeing Android app
-
Built as a group project for university, the application ingests and analyses fitness and calendar data to make recommendations to the user for meals, exercising and sleep. These are dynamically computed to fit in the user’s daily schedule and to optimise his/her wellbeing by referring to NHS advice.
Other creators -
Honors & Awards
-
Distinction in the Honour Moderations in Computer Science
University of Oxford
Awarded by the University of Oxford for academic achievements in the years 2014-2015, 2015-2016, and 2016-2017.
Languages
-
Engleză
Native or bilingual proficiency
-
Germană
Elementary proficiency
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More