Emre’s Tesseract

Welcome in. From the tone of a Turkish sentence to a quantum measurement, these are the questions I keep coming back to.

Can one model edit the weights of another?

I tested edits to a classifier with permanent damage on the UCI Adult dataset. Sol’s edit increased balanced accuracy, but overall accuracy fell and the predicted probabilities became less reliable. For the intact SmolLM2 language model, every selection on ARC-Easy and PIQA kept the original weights.

Damaged classifier

UCI Adult · Sol

BaselineSelected
Balanced accuracy78.59%79.02%
Overall accuracy82.86%82.63%

Intact language model

SmolLM2-1.7B · ARC-Easy / PIQA

Original weights selected

Measurements · GitHub
Emre beside an Elektra espresso machine
My favourite espresso machine.
Emre presenting his PhD research at Navarra
A snapshot from my PhD presentation at the V Jornadas Científicas de DATAI, University of Navarra.
Emre holding the BBVA We Think Big award
I received BBVA’s We Think Big award in 2024 for my work on patents and my academic contributions.

Hi, I’m Emre.

I’m from İzmir, on Turkey’s Aegean coast. I used to be a civil engineer. Then I took an arrow to the knee and traded my hard hat for a lab coat.

I’m an engineer and applied scientist on BBVA’s AI Innovation team. At the University of Navarra, I’m doing a PhD on fairness in AI. My interests range from what makes a good team to abstract maths, quantum physics and philosophy. I especially enjoy finding a link between two ideas that seem to have nothing in common.

Madrid is home now. I’m exploring the city, learning Spanish and getting to know the cultures of Spain and Latin America. At home I paint, play chess and make up little stories for my daughter.

Contact & profiles

Models and code

  • Turkish Sentiment Analysis · TSAM · I trained this BERT model to read the tone of a Turkish review. It scores the sentence for positive, neutral and negative sentiment.
  • Whisper Medium TR · Give it a Turkish recording and it returns a transcript. I adapted Whisper Medium on Mozilla Common Voice 11.0 speech to study recognition errors in Turkish.
  • Gemma 2 · Turkish Wikipedia · This adapter continues a Turkish passage using Gemma 2 9B trained further on Turkish Wikipedia. The card permits academic and research use and limits commercial use to BBVA Group.
  • Java Methods · Doc2Vec · A Java method goes in and a 200-dimensional vector comes out. Developers can use these vectors to find similar snippets or group methods in a codebase.
  • Gemma 3 27B · Turkish reasoning · I fine-tuned Unsloth’s 4-bit build of Gemma 3 27B Instruct for answering reasoning questions in Turkish. The card includes preliminary TARA evaluations on tasks such as logic, planning and SQL.
  • arXiv Abstract-to-Title · mT5 · Paste a research abstract to get a candidate paper title. This mT5-small adaptation learned the task from abstract–title pairs in an arXiv dataset.
  • Wav2Vec2 XLS-R · Turkish · This Turkish speech recogniser turns recorded audio into text using the multilingual XLS-R encoder. Researchers can inspect its training history and its word error rate on Turkish Common Voice.
  • Qwen 0.5B · GRPO maths · A small model for maths word problems, trained on GSM8K to produce a worked response and a numerical answer. The training rewards answer correctness and response format separately.
  • Regression Conformal Prediction Research Atlas · Look up a dataset and a method to see how often its prediction intervals covered the outcome and how wide those intervals were. The atlas links results across 67 datasets to experiments and a browsable knowledge graph.
  • BIRD · Text-to-SQL experiments · The pipeline takes a question and a database schema, proposes SQL and checks the result by executing it. Researchers can inspect how ambiguity, table selection and query validation affect the answer on a financial subset of BIRD.
  • WebCLI · WebCLI indexes public documentation and lets developers or agents search it locally from the command line. Results include stable references and can be exported as JSONL, with OpenAI’s developer documentation as the first adapter.
  • Fairness Metrics · Give the Python library features, predictions, actual outcomes and a protected group. It returns a pandas table comparing quantities such as false-positive, false-negative and recall rates across groups.
  • Quantum Time Experiment · If several fragments of a quantum system preserve the same record, can they tell us what happened earlier? The experiments test which records can be put in order and when an additional reference is needed to determine the direction.
  • PaperBanana field test · I tested Zhu and colleagues’ illustration pipeline on the methods sections of three of my papers. The repository pairs each untouched output with the original figure and exact input, so readers can judge where the illustrations went wrong.
  • World Cup 2026 Nowcast Lab · This research lab turns match probabilities into distributions over possible tournament winners by simulating the 2026 World Cup. Readers can inspect the forecast card, data sources and model choices behind each snapshot.
  • Agent Chronicle · Read a Codex or Claude session as a timeline of decisions, changes and checks. Each recorded event gives the developer a way to inspect the evidence behind a progress report.

Research questions

Fairness

Who carries the uncertainty?

How uncertainty estimates and referral to human review affect different people in decision systems.

Four questions from my notebook

Ontologies and data structuresWhat changes in a graph when a train runs late?

A station is a place. A departure is an event. If I give each its own node type and each time a source, can I trace exactly which connections a single delay affects?

RDF 1.1 · W3C
Meaning-Native ComputingCan a computer predict which instructions will run next if it knows the rules of the program?

Meaning-Native Computing is my independent research project. I want to compare two predictors given the same program and comparable computing resources. One also gets an explicit description of the rules the program must follow. Does that help it predict the next steps? Before a predicted change can take effect, an independent check must confirm that it follows those rules.

Category theoryWhat survives when I compose two transformations?

One process turns a timetable into a graph. Another finds a route through it. When I compose them, how do I preserve which station each time belongs to?

Seven Sketches in Compositionality
Quantum measurementWhat if I ask the same qubit a different question?

I can prepare a qubit so it always gives the same answer when measured along one axis. Along a perpendicular axis, either of two results is equally likely. What fascinates me is how changing the measurement changes the answer while the preparation stays the same.

Quantum Information · IBM

Patent and utility model applications

Public registry and publication records name Emre Taşar as an inventor on 28 patent applications and 2 utility model applications in Türkiye. Filing numbers, titles and dates come from the TÜRKPATENT extract of 2026-08-31, with protection types from the inventor’s list. Current legal status is checked separately in the register. Some inventions also have an international publication.

Applicant · TÜRKİYE GARANTİ BANKASI ANONİM ŞİRKETİ

Original Turkish titles and editorial English translations. TÜRKPATENT register

30 application records

Selected research

Between-group gaps in coverage and referral. Percentage points, 100 random seeds, 90% target coverage.
DatasetCoverage · PooledCoverage · GroupwiseReferral · PooledReferral · Groupwise
Adult5.01.25.112.9
ACS Income0.80.83.45.4
Bank Marketing2.01.10.03.0
COMPAS1.83.35.710.8
German Credit3.55.76.124.7
Taiwan Credit1.11.33.47.7
Preprint, Table 4

Work index

Agent Chronicle

2026 · Public source project · Project maintainer

A local control surface for coding agents that connects progress, decisions, changes, and verification to recorded evidence. Primary source

AI-assisted point of sale

2024 · Published PCT application · Co-inventor

A proposed POS system connecting personalized recommendations, fraud detection, and natural-language device operation. Primary source

Turkish speech recognition

2024 · Published journal article · First author

An evaluation of self-supervised speech representations and transfer learning for Turkish speech recognition. Primary source

Qwen-0.5B-GRPO

2025 · Public model adaptation · Model developer

A GRPO fine-tune of Qwen2.5-0.5B-Instruct for structured mathematical reasoning using GSM8K. Primary source

BLOOM

2022 · Collaborative research paper · Co-author

A contribution to the BigScience collaboration on open multilingual language-model research. Primary source