scott-k.dev
software engineer — AI systems, robotics & infrastructure
Hull, UK
projects
in progress
ViT-Qwen VLM
A from-scratch vision-language model implementing the core architectural primitives behind production VLMs: a dual-encoder vision tower that extracts patch embeddings from raw images, an MLP projector that bridges the visual and language representation spaces, and multimodal token prepending that feeds image patches directly into the language model's embedding stream.
PyTorch SigLIP Qwen3-4B ViT FastAPI
currently: stage-2 diagnostics
read
in progress
Drone RL
Quadcopter hover and target-tracking agent trained with Soft Actor-Critic and Prioritized Experience Replay inside a PyBullet physics simulation. Training metrics streamed live via MLflow.
PyTorch SAC+PER PyBullet MLflow Gymnasium
currently: test
visit teaching a drone to fly
building
System Zero
Self-hosted AI platform. Hybrid semantic search over your own documents, multi-turn reasoning, Text-to-SQL, and a streaming chat interface — all running on local LLMs via llama.cpp.
FastAPI llama.cpp k3s RabbitMQ FAISS React
currently: polish
visit
planned
RLxNEAT
Neuroevolution of Augmenting Topologies combined with reinforcement learning reward signals. Evolves both network topology and weights to solve sparse-reward control problems.
NEAT Reinforcement Learning Python Evolution
currently: dev
planned
writing
Blog
Notes on what I'm currently building, exploring, and thinking about — AI systems, robotics, and the occasional rabbit hole.
AI Robotics Infrastructure Notes
read
lab
>
explore
about
Scott K
Building AI systems and infrastructure that actually run — on your hardware, under your control. Interested in reinforcement learning, robotics, distributed systems, and the full stack from silicon to UI.

Everything here is self-hosted.
github.com/Kikumu
cv
experience
AI Agent Engineer Oct 2020 — Present
  • Built production AI agents with reasoning, planning, and autonomous execution using Azure AI and Anthropic Claude — multi-step tool-calling via MCP with FastApiMCP and explicit operation routing.
  • Engineered hybrid RAG pipeline combining FAISS dense vector search with BM25 sparse retrieval, delivering a 40% improvement in retrieval precision through systematic evaluation.
  • Led end-to-end deployment of 3 concurrent production AI applications — GitHub Actions CI/CD, k3s cluster management, KEDA autoscaling, Prometheus + Grafana observability.
  • Self-hosted GGUF inference via llama.cpp on CPU-only VPS; built a real-time SSE streaming gateway and multi-service platform with FastAPI and RabbitMQ.
  • Applied LoRA/PEFT fine-tuning to adapt foundation models (OpenAI, Anthropic, Llama, Qwen) for domain-specific production tasks; managed token budgets and latency constraints.
technical skills
AI Agents Multi-agent architectures, MCP, tool orchestration, planning/reasoning agents, prompt engineering
RAG & Retrieval FAISS, BM25 hybrid search, semantic chunking, re-ranking, retrieval evaluation, SentenceTransformers
Languages Python (expert), TypeScript, C#, SQL, Bash
Backends FastAPI, FastApiMCP, SSE, JWT Auth, RabbitMQ, MySQL, PostgreSQL
Infrastructure Docker, Kubernetes / k3s, GitHub Actions, Azure (AKS, ACR, AzureML, Key Vault)
ML / RL PyTorch, SAC, TD3, PER, DroQ, PyBullet, LoRA / PEFT, llama.cpp GGUF, MLflow
education & certifications
BSc (Hons) Computer Science, 2:1 2020
University, UK — Honours Project: Autonomous Pathfinding Robot (First Class)
Microsoft Certified: Azure AI Engineer Associate (AI-102) Microsoft Certified: Azure AI Fundamentals (AI-900) | Azure Fundamentals (AZ-900) Modern Reinforcement Learning: Actor-Critic Algorithms (Coursera)