AI systems engineer and UX
Ryan Leibering
I build AI systems end to end, anchored in a knowledge graph the agents reason over, with multi-agent orchestration and hybrid retrieval drawing on sources I curate.
The agents do the work, delegated to them through conversation. They research and design the architecture, store their reasoning as a durable structure, and every new decision is built within it. I call it DARF, the Delegated Architectural Reasoning Framework.
Inside the architecture
DARF turns delegation into grounded, governed work.
When I describe a problem to the agents in plain language, the words enter a knowledge graph rather than only prompting a model to predict an answer. That graph is a dense web of typed, connected facts, holding the research and prior decisions, every one tied to the exact source it came from.
Reasoning runs across that grounded structure, so the few sentences I type set real, source-backed inference in motion. The agents reach a decision and write it back as a typed record that keeps its reasoning. Whatever the agents build is probationary, recorded for me to audit, and advances only at the pace I set.
The design system
The interface builds itself, accessible by construction
The design system behind this site generates the interface rather than only styling it. It is a server-driven UI, which means the page is built from data instead of hand-coded. The layout, the words, and the citations behind them are pulled from the same knowledge graph the agents reason over.
Color is the last thing added, and it can never break the page. Every surface rests on a fixed achromatic scale of brightness steps whose contrast is proven before any palette is applied, so a new theme can repaint the whole interface without dropping below an accessible ratio. The gallery below is the proof. Pick any artwork. The system pulls five colors from it, maps each onto that proven scale, and re-themes the section live while the ratios hold.
Theme: Harbor Keep
The stack
Same discipline, every layer
Every layer of the stack, top to bottom, and the part I built in each.
- 01
Frontend
- Next.js
- Tailwind
- TypeScript
I built the design system everything else is presented through. It can shift its whole look to a new brand or a generated palette while accessibility and compliance stay fixed, so the style changes but nothing it renders ever drops below WCAG 2.2 AA.
- 02
Backend
- Neo4j
- Qdrant
- PostgreSQL
I stood up the data spine the system reasons over. A property graph in Neo4j holds the source of truth, Qdrant answers similarity, and PostgreSQL carries the live task queue the agents coordinate through.
- 03
Agentic harness
- OpenAI
- Claude
- Gemini
- Open LLaMA
I built the harness that lets AI agents claim work, run it at machine speed, and stay accountable for it. Agents from four model providers reach the system through one Model Context Protocol interface, and everything they produce passes the same gates before it can land.
- 04
RAG
- Neo4j
- Qdrant
- BGE-M3
I built the retrieval-augmented generation layer that fuses vector similarity with the graph by reciprocal rank fusion. An answer is assembled from typed nodes and their relationships, so it names exactly what it drew on rather than resting on a bag of matched text.
- 05
Graph ML
- Neo4j
- PyKEEN
- PyTorch
I designed the ontology that structures the whole corpus, the entity and relationship types the graph is built from. On top of it I trained embeddings and a link-prediction model to surface connections no person could enumerate, scored on a held-out split.
- 06
Formal verification
- TLC
- TLA+
- Python
I write the protocols that cannot fail as formal TLA+ specifications, authority and ratification and lifecycle among them, and check every reachable state. When one can break, the model-checker returns the shortest path that breaks it, so the flaw is caught in the design and never reaches production.
- 07
Infrastructure
- GCP
- Linux
- Azure
- Docker
I have run on Azure and GCP, but the system runs predominantly on local hardware I control. The graph, vector, and relational stores and a local model sit in containers on a consumer machine, which keeps the data sovereign and the deployment something anyone could reproduce.
- 08
Generative AI
- TikTok
- Midjourney
- Google FILM
- After Effects
I produce visual music videos for TikTok and other social platforms on a constraint-driven pipeline, fixed-seed Midjourney variations carried into motion by Google FILM and finished in After Effects. Varying one thing at a time is the same habit that later ran the ML research.
- 09
Security
- Foundry
- Solidity
- Code4rena
I find smart-contract vulnerabilities by running graph operations over the code, under a submission discipline I built from my own post-mortems. The misses go on the record next to the wins, so a high-severity bug that paid nothing as a duplicate becomes the rule that catches the next one.
Contact
Let's talk
Open to roles where being correct is the requirement and being clear is the difference: systems engineering, AI infrastructure, and the frontends that make them legible. Email is fastest; I reply within a day.