Projects
Engineering projects by Adam Kaniasty covering mechanistic interpretability, RAG, medical imaging and reinforcement learning, with architecture and evidence links.
These case studies describe my role, the systems I worked on, and the available source code or research. They include open-source projects, university work and professional applications.
Mi-Crow — Mechanistic Interpretability for LLMs
Python library connecting sparse autoencoder training, activation analysis, concept discovery and language-model steering.
Mi-Crow — Mechanistic Interpretability for LLMs source codexLungs — Medical Imaging Inference Application
Java Spring and React application coordinating CT image processing, model inference and radiology reports through RabbitMQ on GPU-powered Kubernetes.
APPI Marketplace — Knowledge Graph RAG
AI agent platform for technical documents, with knowledge graph-based context retrieval, a FastAPI backend and Azure deployment.
Reinforcement Learning for Kubernetes Autoscaling
Nokia engineering work on DDQN reinforcement learning, time-series prediction and configurable training pipelines for Kubernetes pod autoscaling.
APPI Synapse — Document Retrieval Application
Internal RAG application for information retrieval in architecture and construction, built with React, FastAPI, MongoDB, LlamaIndex and LangChain.
RL Doom — Reinforcement Learning in VizDoom
Team project comparing PPO and A2C agents in VizDoom, with image preprocessing, reward functions and TensorBoard evaluation.
RL Doom — Reinforcement Learning in VizDoom source codeCanSat Terrain Image Classification
Python terrain-image processing for Project Trailblazer, with image tiling, a random forest baseline and a transferred ResNet18 classifier.
CanSat Terrain Image Classification source code