Mikil Radu Foss

Loss (BCE, rolling)

Steps: 0 Avg: 0.000 Last: 0.000

Education

University of Minnesota, Twin Cities

Exp. Dec 2025

B.S. Computer Science & Mathematics (Double Major)

GPA: 3.8

Key Coursework: Machine Learning, Harmonic Analysis, PDEs, Quantum Physics, Graph Theory, Operating Systems, Data Structures & Algorithms.

Awards & Competitions

  • UMN National Scholarship ($40,000)
  • AWS DeepRacer ML (4 Top-10, 500–1000 monthly)
  • UNL Math Day (Top 10 / 2000)

Research Experience

Mutual Information Neural Estimators (MINE)

10/23 - 05/25

Mentor: Prof. Andrew Lamperski | Read Paper

  • Established rigorous theoretical guarantees for MINE by deriving quantitative error bounds of $O(m^{-1/2} + T^{-1/3})$ for KL divergence estimation using shallow neural networks.
  • Synthesized advanced techniques from convex optimization and functional analysis to construct the main proof framework.

Oak Ridge National Laboratory

12/24 - Present

Mentors: Dr. Olena Burkovska & Dr. Konstantin Pieper

  • Developing a custom sparsity-promoting algorithm to construct neural networks that approximate the memory kernel in the Generalized Langevin Equation.
  • Achieved a 10x speedup in practice by developing an $O(N)$ SDE algorithmic improvement, replacing the previous $O(N^2)$ approach.
  • Conducted and documented over 100 machine learning experiments to validate the theoretical framework.

Neural Network Compressiveness

02/24 - 02/25

Mentor: Prof. Vladimir Sverak

  • Investigated the compressiveness of neural networks compared to traditional Fourier series, focusing on functions emerging as steady-state solutions to PDEs.
  • Implemented numerical methods such as RK4, Spectral methods, and Godunov's scheme for PDE approximations.
  • Utilized autoencoders and sparse autoencoders to measure and compare compressiveness.

Experience

Creevo.app

Founder

12/24 - Present

  • Platform that generates playable games from natural language descriptions.
  • Built conversational agent pipeline with iterative clarification dialogue; developed Godot engine API wrapper with JSON-formatted responses optimized for LLM consumption.
  • Trained custom Godot coding agent using GRPO with multi-signal grading (compilation, time, LLM-as-judge). Improved compile rate from 80% to 98% and reduced unplayable game rate from 8% to <1%.
  • Built Kubernetes cluster using Google’s agent-sandbox (kubernetes-sigs) for secure multi-tenant agent execution.
  • Led product research, built the product, and drove iterative improvement from customer feedback.
  • Designed the Creevo Agent Library (CAL), an open-source framework supporting multimodal tool responses.
  • Managed a team of 4 researchers and built scaling cloud infrastructure so agents run in Docker with Godot.

Software Contracting

Anchovy

05/25 - Present

  • Built P2P marketplace for recipe sharing for Anchovy in .NET and Angular.
  • Built networking app that securely connects Rockwell PLC machines through mobile phones to a laptop anywhere in the world.
  • Redesigned and refactored a custom OpenAI SDK to support multimodal prompts.
  • Designed and implemented a marketplace feature allowing users to create and sell cookbooks.
  • Led a full app-wide UI/UX redesign for the Ionic Angular frontend.
  • Restructured teams meeting to move the team over to agile.

Amway

Software Engineering Intern

05/23 - 06/24

  • Designed and implemented an AWS system to monitor Kubernetes & Gluu processes.
  • Analyzed user registration data to identify and resolve an issue contributing to a 60% decline in registered users.

Technical Projects & Background

14 Years Game Development — Hobbyist Unity developer since 2012; experience with physics systems, procedural generation, game AI, and asset pipelines.

Creevo Agent Library (CAL)Open-source agent framework supporting multimodal tool responses; integrated with LangSmith and Maxim AI for observability.

Custom ML Library — From-scratch ML framework with transformer support, optimization algorithms (SGD, Adam, Muon), and RL environment integration.

Chess AI — Developed undetectable approximately 2000 Elo engine using classification CNN with adapted minimax search.

Skills

PythonC++JavaJavaScript .NETAngular TensorFlowJAXPyTorchReactJS DockerKubernetesAWSLaTeX LLMsGenerative AIMCP

Side Projects