ML Engineer.

About me

Projects

Sometimes it is the people no one imagines anything of who do the things that no-one can imagine.

Alan Turing

---

Flappy Bird

Flappy Bird is an OpenGL game made as as­sign­ment for a Com­puter Graph­ics course in the Uni­ver­sity. The game is a clone of the pop­u­lar mo­bile game Flappy Bird where the player must tap the screen to make the bird fly and avoid the green tubes. The game was made in C++ us­ing the OpenGL li­brary and the SDL li­brary for the win­dow cre­ation and in­put han­dling. The game fea­tures a main menu, a game over screen, and a score sys­tem. We won the first place in the course com­pe­ti­tion with this game.

One vs One Hun­dred

1 vs 100 for Win­dows Mo­bile was a pro­ject which be­gan as an as­sig­ment of an Uni­ver­sity course in as­so­ci­a­tion with Mi­crosoft Uruguay. The ob­jec­tive was to de­sign a game based in the amer­i­can TV show of the same name. The game must al­low mul­ti­ple play­ers to join a game and com­pete against them­selves re­quir­ing only an in­ter­net con­nec­tion in their phones. We also re­quired to make an Ad­min­is­tra­tion site to man­age the con­tent avail­able within the game. The re­sult was im­pres­sive as we proved that is pos­si­ble to make a qual­ity game in the Win­dows Mo­bile plat­form. The pro­ject was com­pleted in 14 weeks.

RP­Solver

RP­Solver is a tool that was re­leased as part of my pub­li­ca­tion for us­ing Pho­ton Map­ping for solv­ing ar­chi­tec­ture prob­lems. The tools al­lows the user to up­load an scene and the tool will gen­er­ate a pho­ton map ren­der­ing for it. This tool was de­vel­oped in C++ us­ing the Op­tiX li­brary for ray trac­ing and CUDA for the par­al­lel pro­cess­ing.

SCIS

Suit­eCom­merce In­Store is a point of sales ap­pli­ca­tion that lever­age Net­suite ERP ca­pa­bil­i­ties with web tech­nolo­gies. It is de­vel­oped mainly in Javascript and Swift. It pro­vides re­tail­ers with a so­lu­tion that uni­fies the phys­i­cal and dig­i­tal shop­ping ex­pe­ri­ences within a sin­gle, cloud-based com­merce plat­form. Us­ing a mo­bile de­vice, your sales as­so­ci­ates can ac­cess in­ven­tory and cus­tomer in­for­ma­tion to en­gage cus­tomers ef­fec­tively, drive more sales, and pro­vide a sat­is­fy­ing shop­ping ex­pe­ri­ence.

Say Cheese

This pro­ject was the main re­sult of an Uni­ver­sity course I par­ticiped in with 3 pat­ners. The aim of the course was the con­struc­tion of a Web Ap­pli­ca­tion for both events (like par­ties, wed­dings, etc.) hosts and guests to share dif­fer­ent re­sources like pho­tos and videos. The hosts or­ga­nize events on the site and the event guests up­load con­tent from dif­fer­ent places like Youtube, Vimeo or their com­put­ers. In a nut­shell the ap­pli­ca­tion is a suite that al­lows to man­age events and it be­haves like a so­cial net­work in the way that al­lows users to in­ter­act be­tween them by send­ing mes­sages, or things like be­ing tagged in pho­tos. Say Cheese also in­te­grates with Face­book, Twit­ter, Google Maps and al­lows users to safely lo­gin us­ing their Face­book ac­count.

Paste­board

Paste­board is a link shar­ing app ori­ented to share links among cowork­ers, or peo­ple that work in the same pro­ject. It aims to re­place wikis and make shar­ing and main­tain­ing in­for­ma­tion about pro­jects more eas­ily.

Scrap­ing De­vel­op­ment Frame­work

SDF is a Python frame­work that al­lows to easy scrape a site. It en­ables to make XPath and CSS queries to quickly parse a web­site. The fo­cus of the frame­work is to make the job of mak­ing a scraper eas­ier, al­low to parse sites with mil­lions of items, and be able to ex­port the items in sev­eral for­mats rang­ing from CSVs to MySQL data­base dump. SDF uses browsers (as in a Web Browser) to do the job. Browsers come in two fla­vors: the We­bKit­Browser, a stan­dard browser (sim­i­lar to Sa­fari or Chrome) that can be con­trolled through python code. We­bkit­Browser man­ages javascript, flash and other types of con­tent. The other type of browser is the Ba­sicBrowser, a min­i­mal­is­tic and faster ver­sion of We­bKit­Browser that does­n't han­dle javascript di­rectly. SDF fea­tures re­cov­ery and par­al­lel pro­cess­ing, so mul­ti­ple browsers can be used con­cur­rently to scrape a large site in hours.

No­tranzo site

The pur­pose of this pro­ject was to build an ap­pli­ca­tion to sup­port an com­put­ing sal­ing store. The sys­tem must fea­ture a con­trol panel for the em­ploy­ees in dif­fer­ents branches to man­age the prod­ucts, the stock avail­able for each one, make move­ments be­tween branches, and make sales of them. Also a web­site must be build to im­ple­ment a vir­tual on­line branch that of­fers its prod­ucts in the in­ter­net. This pro­ject was made as an as­sig­ment of an Uni­ver­sity course. The pur­pose of this pro­ject was to build an ap­pli­ca­tion to sup­port an com­put­ing sal­ing store. The sys­tem must fea­ture a con­trol panel for the em­ploy­ees in dif­fer­ents branches to man­age the prod­ucts, the stock avail­able for each one, make move­ments be­tween branches, and make sales of them. Also a web­site must be build to im­ple­ment a vir­tual on­line branch that of­fers its prod­ucts in the in­ter­net. This pro­ject was made as an as­sig­ment of an Uni­ver­sity course.

Your Pro­ject?

Feel free to con­tact me if you want to col­lab­o­rate on a pro­ject or if you have any ques­tions. I'm al­ways open to new ideas and op­por­tu­ni­ties. Let's cre­ate some­thing amaz­ing to­gether!

Publications

Somewhere, something incredible is waiting to be known.

Carl Sagan

---

I have cre­ated mod­els and meth­ods to im­prove rec­om­men­da­tion sys­tems and solve light­ing prob­lems in de­sign. My work on Align Macrid­VAE com­bines vi­sual and text data to make rec­om­men­da­tions more ac­cu­rate and eas­ier to un­der­stand. I also used search data and trans­form­ers to learn how items are sim­i­lar, and de­vel­oped a pho­ton trac­ing method to make light­ing de­sign in ar­chi­tec­ture more ef­fi­cient.

• 

  Align Macrid­VAE: Mul­ti­modal Align­ment for Dis­en­tan­gled Rec­om­men­da­tions

  ECIR 2024: Ad­vances in In­for­ma­tion Re­trieval

  Item alignment generated by the recommendation algorithm

  In this pa­per, we in­tro­duce a new rec­om­men­da­tion model that com­bines text and im­age data to sug­gest items to users. By align­ing the vi­sual and tex­tual de­scrip­tions in a shared space, our model bet­ter un­der­stands item fea­tures, im­prov­ing rec­om­men­da­tion ac­cu­racy and help­ing us vi­su­al­ize user pref­er­ences based on dif­fer­ent item as­pects. This pa­per was based on my mas­ter the­sis at KU Leu­ven.

• 

  Learn­ing Ré­sumé Em­bed­dings with Search Data and Trans­form­ers

  2023 IEEE In­ter­na­tional Con­fer­ence on Sys­tems, Man, and Cy­ber­net­ics (SMC)

  Model architecture for learning résumé embeddings

  In this pa­per, we ex­plore how users' clicks and in­ter­ac­tions with search re­sults can teach us about the hid­den con­nec­tions be­tween dif­fer­ent search re­sults. By an­a­lyz­ing these in­ter­ac­tions, we can build a model that iden­ti­fies which search re­sults are sim­i­lar in a way that is­n't ob­vi­ous at first. We use ad­vanced tech­niques like con­trastive learn­ing with BERT mod­els to train this sim­i­lar­ity model, which helps us make bet­ter rec­om­men­da­tions based on what users are ac­tu­ally in­ter­ested in. This work was based on re­search made while I was work­ing at Hired.

• 

  A Pho­ton Trac­ing Ap­proach to Solve In­verse Ren­der­ing Prob­lems

  Con­fer­ence On Graph­ics, Pat­terns And Im­ages (SIB­GRAPI), 2018

  Conference solution generated by the optimization algorithm

  In this pa­per, we in­tro­duce a new method to help de­sign­ers achieve their light­ing goals more ef­fi­ciently. By com­bin­ing pho­ton trac­ing with an op­ti­miza­tion tech­nique, our ap­proach han­dles var­i­ous light­ing needs with­out re­quir­ing lots of com­put­ing power, and it works well even when the space's shape is part of the de­sign. This pa­per was based on my Com­puter En­gi­neer­ing the­sis at Ude­laR.