Aurorium NexusPost-trained models, RAG pipelines and agentic systems.
Applied AI engineer. Below is what I've built, why I built it, and where to find it.
Projects
All repositories-
Distillery
- What
- A sandbox-verified distillation autopilot on Nebius Token Factory: it trains a small model to take over a narrow job from a big one, behind an OpenAI-compatible gateway.
- Why
- Big models are overkill for narrow, repetitive jobs. I wanted a small specialist that takes over only once it is proven as good, and hands back anything it can't guarantee.
-
Post-training Skills
- What
- A field guide and local deep agent for learning the major parts of LLM post-training.
- Why
- To map all of LLM post-training in one place, from data and SFT to preference methods, RL and evaluation, for myself and anyone else learning it.
-
Aurorium Mind
- What
- A QLoRA-tuned Qwen 3.5 4B served with vLLM, Rust and FastAPI, with a Next.js workspace on top.
- Why
- To take a fine-tune past the notebook: train it, serve it on my own stack, and put a real product in front of it.
-
Canary Backend
- What
- The backend of a release gate that red-teams AI agents over HTTP and replays the attacks against a baseline, returning PASS, WARN or BLOCK. I contributed 25 commits.
- Why
- To turn red-teaming into a release gate, so a team gets a clear ship-or-block answer before an agent goes out.
-
Agent Canary
- What
- A red-team workflow for testing HTTP AI agents and validating the mitigations.
- Why
- To attack AI agents for real: live targets, real responses and saved evidence, with no mocked results.
-
AudioForge
- What
- A reproducible audio-ML training pipeline with LoRA adaptation and published artifacts.
- Why
- To compare audio models under one explicit pipeline instead of hiding the experiment inside a notebook.
-
IncidentOps
- What
- A collector-first incident investigation system: engineering evidence in, cited and reviewable findings out.
- Why
- A repository alone can't explain an outage, and tools that guess at root causes are risky. I wanted answers that cite their evidence and say what is missing.
-
LogSage
- What
- A QLoRA-trained model that turns application logs into structured diagnoses.
- Why
- To learn the full fine-tuning loop end to end: dataset design, QLoRA training on real AWS GPUs, observability and publishing.
GitHub, last 18 months
github.com/Auro-rium—public contributions
TensorTonic
tensortonic.com/profile/aurorium-nexusWorking through machine-learning problems from first principles. This is my public profile card, straight from TensorTonic.
Open profileOpen to applied AI roles and hard problems.
Internships, collaborations, or anything where the model has to actually work in production.