Nikos Fourlis

Second person on the team, backed by a16z Speedrun, Entrepreneur First, OpenAI, and Berkeley SkyDeck. I work across the extraction pipeline, backend, and internal tooling, and directly with early customers. I own the confidence layer that decides which outputs file automatically and which go to human review.

Part of a small team building the full production stack for Sophea AI, Greece's first sovereign large language model. Work spans model training on NVIDIA DGX B200, voice stack integration (ASR/TTS), and multimodal capabilities. The model serves regulatory document analysis, accounting automation, and enterprise customer support.
Cut a vision-language model's GPU memory 3.6× (18 GB to 5 GB) with FP8/INT4 quantization, then shipped it as a vLLM serving stack with an accuracy/latency benchmark suite (GitHub). Ran a 12-model training study predicting how performance scales with model size and data: results matched OpenAI's published scaling laws, and showed more data beats a bigger model at equal compute (GitHub).

Collaborated on the architectural design and prompt engineering of Aktor's first internal AI Agent. Optimized enterprise reporting pipelines by structuring large-scale SAP ERP datasets via complex SQL queries, bridging AI tooling with legacy enterprise infrastructure.

One of Europe's most rigorous engineering programs. Five-year integrated master's covering electrical engineering, computer science, applied mathematics, AI/ML, and robotics. GPA: 8.62 / 10, taking master's-level coursework from year 4.

Interpretability, machine unlearning, and model introspection: whether a model's self-report tracks its internal recognition signal, and whether it can tell suppression apart from genuine removal from the inside.

Suppressed knowledge in language models. Fine-tuned Qwen3.5-4B to refuse 53 facts it knew, then probed whether the knowledge survived. The fine-tune installs a general refusal, not a fact-specific gate, and no selective checkpoint exists across six training levers. Linear probes on frozen base activations still recover the answer at the subject token 75–80% of the time, which explains why 15 inference-time elicitation attempts failed.

A controlled study of catastrophic forgetting in small language models. Trained a 17M parameter GPT from scratch on 1B tokens of standard English, then fine-tuned it on Gen Z slang across eight data regimes, measuring WikiText-103 perplexity and style shift at every step. Key finding: perplexity increases 47x after just 1k tokens of fine-tuning, while style acquisition plateaus at 18–20% regardless of data volume. Ships with a custom C++ inference engine compiled to WASM, with int8 quantization, a KV cache, and SIMD intrinsics.

Real-time karate technique scoring from a single phone camera, benchmarked against national-level execution. Built during the Kiefer/NVIDIA Open Hackathon on B200.
The pipeline extracts COCO-17 skeletons with YOLO11n-pose, aligns them to champion references via multi-reference Dynamic Time Warping, and scores technique quality with a 252K-parameter ST-GCN. A WKF-inspired biomechanics layer feeds a chain-of-thought coaching system that cites specific WKF rulebook articles on every tip. Pearson r = 0.836 against expert human scoring through over 100 validation runs. 68.7% MSE reduction over baseline. TensorRT fp16 inference: 8.81× speedup.

A single-screen browser game: a karateka punching, kicking, and dashing through endless zombie waves. Built with Next.js 14 and TypeScript, rendered entirely client-side on an HTML canvas with no assets, backend, or game engine. Features a day/night cycle that ramps up the pressure, combo chains, heart drops, telegraphed boss fights, and a localStorage best score. Playable right here on the site.

Scalar-valued autograd engine and neural network library from scratch in C++. Implements dynamic computational graphs, low-level memory management, and full forward/backward passes without any external ML dependencies. The foundation of all subsequent deep learning work.

Conversational AI assistant with persistent memory, deployed on WhatsApp. Designed and deployed using Anthropic APIs. Implemented backend logic for user context tracking, session memory, and automated response handling. First end-to-end deployed AI product.

Led a 5-person team through a 4-day €10K hackathon organized by Kiefer, NVIDIA, and OpenACC. Built Miru, an AI karate coaching system, from idea to live demo on NVIDIA DGX B200. Featured in Kathimerini.

Hand-picked out of 30,000+ applicants for YC's flagship in-person program in San Francisco. Under 5% acceptance rate.

Selected for Philoxenia batch 005, a San Francisco builder program by Hustle Hours, out of 1,000+ applicants (~2% acceptance).

First place among 37 participants in an intensive LLM program organized by Kiefer, NVIDIA, and OpenACC. Applied Test-Time Scaling and fine-tuning to build an AI-powered chat summarization system for enterprise customer care. Led directly to being hired as a junior ML engineer the following week.

Distinguished member of the Hellenic Karate Federation's Pre-National Team squad. Competing at Greek national qualifiers. Daily training maintained alongside a full engineering degree and a professional role.

Active member. Participating in ML and software engineering workshops, AI sessions, and student-led technical projects.