Mihail-Dragos Velicu

Research Interests

My work has one thread through it: inferring something unobservable from an indirect signal, and then proving the inference can be trusted. Cortical sources from scalp EEG, understanding from a learner's answers, expertise from the traces a person leaves while working. That thread runs through real-world evaluation, tacit expertise, and the memory and retrieval layer underneath both.

Research Experience

KnowledgeCore: LLM-as-Judge Evaluation for an Intelligent Tutoring System (Bachelor's Thesis) - Romanian-American University, School of Computer Science for Business Management, Bucharest, Romania
  • Built a system that had to judge the quality of a learning process, not just the final answer: an LLM-as-judge evaluator on a six-dimension rubric, with agreement against human raters measured, not assumed, on the web-search component: Cohen's kappa 0.21/0.31 (provisional; the six-dimension rubric judge's agreement with human raters was not separately measured)
  • Deterministic, mastery-based adaptation layer on top of the judge, so the model's score drives a rule, not a vibe
  • Ablations: a four-engine retrieval comparison and an embedding comparison
  • Cost telemetry: approximately $0.033 marginal cost per learning journey
  • Submitted to BRAIT 2026 (Best Romanian AI Thesis Awards, double-blind track); outcome pending. Manuscript available on request
Junior AI Researcher, Center for AI Research (CRAI) - Romanian-American University, Bucharest, Romania
  • NGIS (NeuroGraph Inverse Solver): a graph attention network with leaky integrate-and-fire spiking neurons, reconstructing cortical activity from 128-channel EEG (OpenNeuro ds003766, 31 subjects)
  • Trained on 4x NVIDIA GH200 GPUs via PyTorch DDP (approximately 17 minutes), built in about 48 hours by a 3-person team at the EuroCC/OpenACC AI-for-Science Hackathon
  • RMSE 0.258 +/- 0.248 against a CNN baseline of 0.185 +/- 0.167; the baseline won. Unpublished MDPI-style manuscript co-authored with Prof. Andrei Luchici
  • Built 2-class and 4-class motor-imagery EEG classifiers on 128-channel data (ICA, band-power features, CNNs, MNE-Python)
  • Replication and evidence review of published AI claims, including tests of AI self-replication claims, with Prof. Andrei Luchici

Selected Projects

Drone Object Tracker - independent, github.com/m-dragosvelicu/drone-object-tracker
  • Real-time YOLOv8 object tracking on a DJI Tello drone, with an optional BrainBit EEG layer gating takeoff via blink detection and automatic landing on inactivity or signal loss: physical-world AI and multimodal signal fusion in one artifact
BrainBit EEG Toolkit (Python) - independent / CRAI-adjacent, github.com/m-dragosvelicu/drone-object-tracker (bci-brainbit module)
  • Open-source Python toolkit for the BrainBit EEG headband: BLE discovery/connection, contact-quality checks, circular-buffer streaming, and a real-time blink detector; instrument-level data capture, a smaller-scale version of a collection problem
YourHealth - university project, Romanian-American University, Bucharest, Romania
  • Lab-report PDF ingestion via a multimodal Gemini 2.5 Pro call: unstructured documents into structured, time-series health records. Dominant contributor (201/242 commits, approximately 83%) on a four-person team (Next.js, tRPC, Prisma/Postgres)

Industry Experience

AI/ML Engineer - dConnection (freeme/dconnection.at), delivered through Sapio AI Studio, Bucharest, Romania (remote team)
  • Built production components of FreeMe, a multi-channel AI assistant (WhatsApp, Gmail, Calendar, SMS, voicemail), within an approximately 30-person team
  • Delivered the "Context Machine" memory service: long-horizon state kept and retrieved across a process history (ingestion, segmentation, intent/contact resolution, retrieval via Qdrant/Redis)
  • Prototyped privacy-preserving embedding transforms for multi-tenant retrieval; delivered an LLM-based urgency classifier with structured reasoning output
AI Software Engineer - Sapio Artificial Intelligence Systems SRL (Sapio AI Studio), Bucharest, Romania
  • Joined a 4-person core team as an intern (Mar - May 2025), shipping prototype-to-production on a lean-scrum stack; retained as AI Software Engineer
  • Built LLM test-case generation (Adapta Robotics) and constitutional-AI urgency classification
Founder - NeoCore Horizons (own software studio), Bucharest, Romania
  • Built CarbonX, an IPCC Tier-1 emissions calculator, running a live pilot with InVivo Group (France)

Teaching and Mentoring

Mentor and Technical Lead, ASTRA 2025 Summer School - ASTRA (summer school, RAU/CRAI-affiliated), Romania
  • Designed a scaffolded two-week BCI curriculum (neuroscience fundamentals, signal processing, Arduino/servo hardware integration); mentored two groups to ship working EEG-controlled prototypes on a 4-electrode headband. Direct observation of where novices fail on a physical protocol that the written protocol does not capture
Introduction to BCI (online course) - RAU Academy / CRAI, Bucharest, Romania (online)
  • Built a free 4-module online course on EEG analysis and modeling, with matching instructor notebooks, scripts, and slide decks
Private Tutor, Python, Algorithms and Machine Learning - Independent (student at IT University of Copenhagen), remote / Copenhagen, Denmark
  • Tutored an IT University of Copenhagen student one-on-one, from introductory Python and problem-solving through full algorithms and data structures, then coached a skin-lesion image-classification ML project

Awards

LUMEN, Council of Europe Democracy Hackathon (Challenge 2: Coordinated Inauthentic Behaviour) - Council of Europe, Strasbourg, France
  • Led a 3-person team building LUMEN, an open auditable observatory for coordinated inauthentic behaviour in online hate narratives; selected from 140+ applicants across four challenge tracks
  • Won the track's Lion's Den stage and reached the cross-track Dragon's Den final, securing a $10,000 Microsoft Azure grant
Valedictorian, Class of 2026 - Romanian-American University, Bucharest, Romania
  • Ranked top of cohort in the B.Sc. Computer Science for Business Management program; selected to deliver the graduation address
2x First Prize (Premiul I), RAU Student Scientific Communications Session - Romanian-American University, Bucharest, Romania
  • "Your AI Friend" (Dec 2023) and "Destupefyinator: How Algorithms Wreck Our Brains, and How We Take Back Control" (Apr 2025), both first prize in the Business, Economics and IT section

Talks and Media

  • "Brain-Computer Interfaces x AI," AIIS 2025 Conference, Bucharest (Jun 2025); "The Fundamentalist: AI Thinking," a 70-minute lecture at ASTRA 2026 Summer School (Jul 2026); breakthrough pitch, Falling Walls Lab Romania (May 2025)

Education

B.Sc. Computer Science for Business Management - Romanian-American University (RAU), School of Computer Science for Business Management, Bucharest, Romania
  • GPA 9.63/10. Valedictorian, Class of 2026 (graduation address). Member, Center for AI Research (CRAI)
Erasmus Exchange, Cybersecurity, AI, Business Intelligence - VIVES University of Applied Sciences, Kortrijk, Belgium
  • One-semester exchange; Linux/AI/Python track building iterative real-time AI application projects
Systematic Trading Course - Imperial College London, London, UK
  • ML for trading: return-prediction models, limit-order-book classification, execution backtests, supervised learning on high-frequency time series

Technical Skills

Research stack: PyTorch (PyTorch Geometric, spiking networks), distributed training (DDP, SLURM, multi-GPU), MNE-Python and EEG signal processing, scikit-learn, retrieval and RAG (LangChain, Qdrant, ChromaDB), LLM-as-judge evaluation design, OpenCV
Programming: Python, TypeScript/JavaScript, Kotlin, SQL
Web/Backend: FastAPI, Next.js/React, Docker, PostgreSQL/Prisma, Redis
Languages: Romanian (native); English (C2, Cambridge English C2 Proficiency certificate)