I’m Altay, a design-focused engineer and creative based in London, with roots in Istanbul. I’ve spent the last few years building software for startups and R&D teams, with a foundation in computer vision, deep learning, generative AI, and large language models.
Away from the terminal, I design posters, host events, and DJ. I care about aesthetics in everything I make, not just how it looks but how it feels to use, and I try to reflect that taste across everything I do. I treat design and engineering as one and the same craft, so what I build is as fast and solid as it is considered.
Product Engineer (part-time), Kahuna Music | Oct 2025 to present
I own two products end to end. One is an internal artist scouting tool built on Kahuna's own revenue prediction ML. The other is a funding product that artists, managers, labels and our team all use. The ML is purely predictive and the customers are creative people, so how the product feels matters as much as how robust the backend is.
Software Engineer, Siemens Advanta R&D | Nov 2023 to Sept 2025
I built smart building apps on the Building X platform, pulling real-time air quality and occupancy data from 80+ sensors through AWS and handling a 3x jump in connected sensors with no slowdown. I also led the integration of ML models for predictive analytics across 21 office buildings (15% lower projected energy use), and built the data pipeline behind automated HVAC adjustments (30% less manual intervention).
Software Developer and Associate Product Manager, Kahuna Music | Mar 2023 to Nov 2023
Core engineer in a small founding team. I owned the full stack (React, TypeScript, Django) and led the revenue prediction engine that cut manual royalty and valuation work by 50%.
MSc Artificial Intelligence for Sustainable Development, UCL (2025 to 2026). Dissertation on memory security in persistent LLM agents, with Holistic AI. Won the Responsible AI red-teaming hackathon. Runway AI Student Ambassador.
BSc Computer Engineering, Boğaziçi University Istanbul (2018 to 2023), ski team co-captain
Programming: Python, JavaScript, TypeScript, C++, Java, SQL
ML & AI: LLMs, Evaluation, PyTorch, TensorFlow, Scikit-learn, Pandas, Jupyter, AWS SageMaker
Cloud & DevOps: AWS (EC2, Lambda, S3), Docker, GitHub Actions
Frontend: React, Angular, Node.js, HTML/CSS
Other: Django, Figma, Solidity
Languages: Turkish (native), English, German
An observability tool for agent memory. See what an agent stored, retrieved and used, and trace any item back to the call that created it.
Agents now carry memory across sessions, but there's no common way to see what that memory did. Persival breaks an agent's run into its individual model calls and reconstructs the memory operations alongside them: what was written, read or updated, which call did it, and which stored items it touched. Every item links back to the call that produced it and the calls that later used it, so you can trace a fact across sessions. It works across memory systems, and it can replay adversarial attacks and flag which actions were harmed through the agent's own memory.
Stack: Next.js, React, TypeScript, React Flow, Dagre, Tailwind CSS, JSON Schema, Python
https://huggingface.co/spaces/holistic-ai/persival-observability
MemGPT-style memory for OpenClaw agents. Public plugin, nearly 2,000 downloads.
OpenClaw had plugins for Mem0 and MemOS but not MemGPT, so I rebuilt MemGPT from the paper as a plugin. Agents get tiered core, archival and recall memory, manage it themselves through tool calls, get memory pressure warnings, and use recursive summarisation. It ships with an install wizard, and it's still maintained (v1.3.2).
Stack: TypeScript, Python sidecar, npm, PyPI
https://www.npmjs.com/package/openclaw-memgpt
https://pypi.org/project/openclaw-memgpt-sidecar/
Attention U-Net that maps deforestation after wildfires in Mediterranean Turkey.
An implementation of an attention-based U-Net, following John and Zhang's paper, applied to deforestation caused by recent wildfires in Mediterranean Turkey. It covers training and evaluation end to end. It's trained on Copernicus Sentinel-2 imagery from before and after the 2021 Manavgat fire in Antalya, and reaches a Dice score of 0.39 (87% accuracy) on the test set; most of the gap to the original Amazon model comes from noisy ground-truth labels, not the change of region.
Stack: Python, TensorFlow, Jupyter
https://github.com/xltvy/attention-unet-wildfire-deforestation-turkey
Reproducible PyTorch benchmark comparing CNN and residual architectures on chest X-rays.
A controlled comparison of CNN and residual architectures for chest X-ray classification, with fixed augmentation and learning rate scheduling so the differences come from the models and not the setup. Built to be reproducible. On the Kaggle chest X-ray (pneumonia) dataset, the ResNet-style model with augmentation and cosine scheduling did best: 95.67% accuracy on the held-out test set.
Stack: Python, PyTorch, Jupyter
https://github.com/xltvy/xray-classification-benchmark
A creative studio's website built as a fully interactive desktop OS. Designed as a team, coded, deployed and content-managed by me.
Studio Wino is an award-winning creative studio in Istanbul, and its website isn't a page, it's a desktop. It has its own file system with folders and files you can open, desktop menu items, an animated background you switch on with a toggle, and an embedded media player. I built it by hand from scratch in React, with no template or site builder. I also handled deployment and content on Cloudflare, so new work goes live without a code change.
Stack: React, Cloudflare
I design stuff. Mostly event posters. But isn't everything in life a design choice anyway?
Dealing out grooves at night. Mostly house and tech house. UCLEMS alum. I've played at Under Istanbul, G Pera Istanbul and Electrowerkz London, and on Brixton Radio. Love playing for friends and family, and nothing beats seeing everyone lose themselves to the tunes. Didn't record all my sets (new year's resolution: record more), but the ones I did are up on my SoundCloud. Give them a spin. Happy to play anytime, anywhere.
Jam Sessions is a series of house parties we host, where we play all night long, gather friends, meet new people and enjoy the music. For each volume, I design a poster and a sticker.
Citadel's 2813 Brewery
From 2021 to 2022, I experimented with brewing my own beer at home: sourcing hops from around the world, trying different recipes, naming each batch after its beer type, and designing a unique sticker and branding for each one based on the inspiration behind it. It was a lot of fun back then, and my friends loved the beers and kept asking for more. Maybe I'll come back to it in the future. I still remember the taste. Such a good icebreaker, too...