This page is the long version. The CV is the short one.
Factored evaluation at Ought
My first paid research work. Ought is an alignment research lab, and this was their empirical programme on scalable oversight, in the same lineage as AI safety via debate. The protocol had three roles. An honest expert writes a faithful summary of an article and answers questions truthfully. A malicious expert writes a misleading summary and answers deceptively. A judge, who never sees the article, has to decide which summary is honest using five top-level questions, each answerable with follow-ups up to five levels deep.
I played all three roles. I also compiled strategy guides and briefing documents for the other experts and judges, which nobody asked me to do.The malicious role taught me more about verification than the judge role did. You learn what a convincing lie needs: a true-sounding frame, quotes that are real but cropped, and a judge with no time.
Human-feedback data for OpenAI language models
Through the Ought connection, Long Ouyang, a researcher at OpenAI, contracted me on Upwork to produce human-feedback data for language model training. That contract ran for five years. It started as summarization review and grew alongside the methodology itself into full RLHF.To be precise about the arrangement: this was freelance work contracted by an OpenAI researcher, logged on Upwork. I was never employed by OpenAI.
The work, roughly in the order it arrived:
- Adversarial testing. Writing code with deliberately subtle logic bugs, to see where model reasoning and error detection break down.
- Expert grading. Postgraduate-level physics problems, programming challenges, Python and C# implementations.
- Long-form evaluation. Deep-research style outputs judged against granular criteria, with a written rationale for every judgement.
- Long-context data. Book-length summarization in the 4k to 8k token era, around 2023.
The tasks kept getting harder, so I kept studying, including formal coursework in programming and algorithms. Then, around the arrival of reasoning models, the outputs changed character. They held together across thousands of words. I began spending serious time catching mistakes in physics domains I had never studied, and I could feel the point where I stopped being a sufficient check.This is the same problem I had watched from the judge's chair at Ought, five years earlier, now at a scale where it stopped being an experiment.
Frontier-model training at Turing
AI trainer and task reviewer for STEM in Turing's frontier-model training programme. Promoted to task reviewer in my first week; the delivery manager later flagged me as a top performer. I authored postgraduate-level physics questions and solutions for RL training, designed agentic benchmark tasks, and reviewed STEM training data for correctness, rigour, and rubric adherence.
AI advisor at Dineline
Dineline is a digital marketing company serving restaurants. I built three separate systems, each replacing work that would otherwise have meant hiring. A prompt-chain tree for ad copy across Google, Facebook, Instagram, and email, structured so the team could re-run and fork branches rather than start over. An employee onboarding assistant tuned on Dineline's own material. And a 40-page book draft on restaurant marketing, produced in days rather than weeks. Over that period the ad spend they managed for customers grew from roughly $300K to $1M a month, with no new hires.The common thread was deciding what needed a model, what needed ordinary automation, and what just needed rethinking. Most of the value came from the third category.
Elicit QA at Ought
A separate engagement from the 2019 research. I tested Elicit, Ought's AI research assistant, for eighteen months, contributed to the testing guidelines, and reported dozens of defects.
Market research at IN Consulting
Market reports on cloud computing, SD-WAN, CCaaS, and Six Sigma training programmes.
Production work on AI safety and policy papers
I did LaTeX production and typesetting for a cluster of AI safety and governance publications, working directly with their authors. This is production support, not authorship. What it gave me was the working relationships and a close reading of the work.
- GPT-4 System Card, with Sandhini Agarwal
- Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims, with Gretchen Krueger
- Computing Power and the Governance of Artificial Intelligence, with Girish Sastry
- Generative Language Models and Automated Influence Operations, with Girish Sastry
- Personhood Credentials, with Zoe Hitzig
- Frontier AI Auditing, with Miles Brundage
- Confidence-Building Measures for Artificial Intelligence: Workshop Proceedings
- Suitability for Large Language Models (sllm), with Sam Manning
- The GPT Dilemma, with Alan Hickey
Education
M.Sc. in Physics, Middle East Technical University, 2016 to 2019. Thesis: Using Artificial Neural Network Techniques for Solar Irradiation Predictions. An applied machine-learning study benchmarking neural networks against classical empirical models.
I engineered a dataset of about 80,000 daily observations from 46 stations across 39 cities, out of inconsistent meteorological records, and translated solar geometry into features: declination, sunset hour angle, extraterrestrial radiation. I tuned Keras networks with Hyperband and ran 50 independent runs per condition to separate real effects from initialization noise. Train, validation, and test splits were disjoint by station, so held-out locations measured spatial transfer.Partway through I found that an earlier set of splits overlapped. I marked the affected results as faulty in the record, kept them, and re-ran the experiments. The 1,600 or so serialized model artifacts stayed labelled as raw, superseded, faulty, or final so anyone could audit which was which.
The conclusion was that the networks performed about as well as classical regression, with the gap inside instrument accuracy. A null-ish result, reported as one.
I was a teaching assistant for first-year undergraduates. Coursework included Deep Learning and Computational Materials Science.
B.Sc. in Physics, METU, 2010 to 2016. Solar energy and computational physics. Thesis on residential space heating performance of solar collectors in Ankara using the f-chart method.
Computers and Programming, Hacettepe University, 2006 to 2009.
Publication
E. C. Akbaba, E. Yüce and B. G. Akınoğlu, "Deep Learning Algorithm Applied to Daily Solar Irradiation Estimations," 2018 6th International Renewable and Sustainable Energy Conference (IRSEC), Rabat, Morocco, 2018, pp. 1–4. doi:10.1109/IRSEC.2018.8702963
Skills, plainly
AI and LLM. Evaluation and rubric design, RLHF and human-feedback data, reviewer calibration, prompt engineering and chaining, adversarial testing and red teaming, agent scaffolding and tool design, structured output, context management, AI-assisted development workflows.
Machine learning. Neural networks and regression, hyperparameter optimization, experimental design, repeated-run analysis, spatial validation and leakage prevention, model benchmarking.
Programming and data. Python (NumPy, Pandas, Keras, scikit-learn, SciPy, Matplotlib), LaTeX, schema design, normalization across inconsistent sources, multilingual datasets. I read C# and build Flutter through AI-assisted development.For the last year and a half nearly all of my code has been written with AI assistance. My hands-on coding is rusty and I am not current on PyTorch. My edge is judgement about model behaviour, evaluation design, and system architecture.
Languages. Turkish (native), English (fluent, working language for over six years).