Master’s thesis researching real-time anomaly detection on streaming sensor data from commercial vehicle fleets. Developing online learning methods that run on edge hardware and share knowledge across the fleet via federated learning — without centralizing raw data. At Scania CV AB, supervised by Dr. Juan Carlos Andresen.

Pre-launch MVP. Traditional productivity apps require the executive function they’re meant to help with. Unspool removes that — you brain-dump in natural language, the AI organizes invisibly.

• Chat-only interface: no lists, no categories, no priority scores
• Suggests one thing at a time based on urgency and your energy level
• Remembers context across weeks and learns what works for you
• Append-only event stream as single source of truth — all views derived, nothing categorized upfront

Stack-based programming language with a bytecode compiler, Cranelift JIT, and GPU backend. Written in Rust. Designed so the compiler acts as a reward function for LLM code generation — small action space (~140 tokens vs ~50K for Python), deterministic semantics, and capability-gated security. Fine-tuned DeepSeek-R1-14B achieved 99.6% accuracy across 478 tasks.

Declarative Python framework for building AI agents using YAML specs. Provider-agnostic, supports parallel node execution, extensible via a registry system for custom tools and LLM adapters. Start simple, scale to advanced parallel workflows.

Workflow engine that runs YAML-defined workflows with P2P tool discovery. Written in Rust. Tools can be builtin, external processes, WASM modules, or discovered from peers on the network via iroh P2P.

Graduate project supervised by Prof. S. Toor & Prof. P. Singh. Conducted systematic empirical study of different type of MoE architectures on CIFAR-10.

Built a custom Transformer architecture from scratch in PyTorch to deepen understanding of attention mechanisms and sequence modeling. Implemented multi-head attention, positional encoding, and encoder-decoder architecture with detailed documentation of design choices.

Developed a distributed text summarization system for large-scale document processing. Implemented parallel processing architecture with task distribution, efficient batching, and result aggregation for high-throughput summarization workloads.

User-friendly apartment search with natural language queries. Built preference parsing system, intelligent ranking algorithms, and application guidance features to simplify the housing search experience.