Portfolio · LinkedIn · me@mattschramm.com
The context an agent needs, the memory it retrieves, the checks that prove it worked, and the product a person can use.
I'm Matthew, a Graduate AI Engineer at Ubundi in Cape Town. I work across product, engineering, and deployment. I take an unclear idea to working software, then stay with it through releases, customer setup, and daily use.
Ayda — company knowledge that agents can use.
My main focus. Ayda is an AI assistant that answers from a company's own knowledge, cites its sources, and respects each person's access. I manage releases and hosting for six company deployments, build connectors, and help customers get to their first useful answer.
Agents and memory — context that lasts beyond a chat.
I built Cortex, a long-term memory system for agents. It combines structured facts, a knowledge graph, and several retrieval methods. I also build deployment tools and integrations, and I keep company agents running. I run my own agents every day; that is how I learn where these systems fail.
First Motive — robot demonstrations into usable datasets.
I built data processing, quality checks, annotation, and storage workflows for robot and human recordings. Now I process and improve robot demonstration data: review, provenance, and checked datasets for the team's training workflow.
| Project | What it does | Built with |
|---|---|---|
| Atrium | A macOS workspace for meetings, notes, and company context. On-device transcription and dictation, with explicit control of each external AI request. | macOS · local models |
| ProjectForge | A CLI that turns requirements and team conventions into verified Python and TypeScript projects through coding agents. | Python · Claude Code · Codex |
| Umbono | Run one prompt across many models. Compare latency, tokens, and cost, and rank the outputs with a weighted human rubric. | TypeScript |
| Resonate | Models a person's communication style, transforms text, and shows each correction. Deterministic demo with synthetic data. | TypeScript |
| Personal Codex | A personal RAG interface with response modes, document workflows, and source attribution. Deterministic demo with synthetic data. | TypeScript |
| Local Context Engine | Chat with your PDFs fully offline through Llama 3.2. A feedback loop records each hallucination. | Python · Ollama · Streamlit |
| MyAdvisor | My UCT capstone, built with two teammates: a role-based academic advising and booking platform. | Java · Spring Boot · MySQL |
- Use AI fully. Keep human judgment. I use Claude Code, Codex, and my own agents to explore and build. The result stays my responsibility.
- Make the evidence visible. Sources, evaluation results, data provenance, and repeatable checks belong in the workflow.
- Follow through. A feature also needs a release, a usable interface, and a person who can operate it.
- Build: Python, TypeScript, React / Next.js, PostgreSQL, APIs
- AI: agents, MCP, retrieval, memory, evaluations, OpenAI, Anthropic, AWS Bedrock
- Ship: AWS, Terraform, Docker, CI/CD, OAuth integrations
- Robot data: ROS 2, MCAP, LeRobot datasets, SmolVLA data preparation
I studied Computer Science and Information Systems at UCT, then completed Information Systems Honours with distinction. My thesis examined why young people share data online when they have privacy concerns.
I also tutor as a volunteer in Langa, where I teach computer skills and beginner coding. When I can explain a subject clearly, I know that I understand it.
My earlier student projects are still here. They show where the work started.
Do you work on agent memory, context systems, evaluations, or robot data?
I'd like to compare notes. Email me or see more of my work.