AI & machine learning engineering student in Stockholm — looking for an LIA placement.
I build LLM systems and the tooling around them: agents that use tools safely, guardrails that get tested against real attacks, and developer tools that make what a model is doing visible. Most of what I build starts because I wanted the thing to exist.
| Period | Dates |
|---|---|
| LIA 1 | 26 Oct 2026 – 1 Jan 2027 |
| LIA 2 | 11 Jan – 2 Apr 2027 (degree project running in parallel) |
Studying AI and Machine Learning Engineering at Teknikhögskolan Stockholm (higher vocational education, HT25–VT27). Based in Stockholm.
Agent RAG (private, 5-person team) — a routing agent that directs a question to the right RAG database among several. I own the retrieval layer and the evaluation: the strategy contract and its keyword/semantic implementations, the query log, and the pipeline that scores four ranking strategies against a ground-truth set.
I wrote the evaluation protocol before any result existed, with a revision log recording whether each threshold changed before or after the number it governs had been read. When the measurement came in, the headroom gap cleared my pre-registered threshold — but the bootstrap interval included zero and the permutation null produced a larger gap on average, so I reported it as no effect. Scrum master for one sprint; 22 % of 603 commits.
| Project | What it is |
|---|---|
| llm-guardrails | Three layers of prompt-injection defence in front of local LLMs, measured against 22 attacks. All layers together reached 0 % attack success — then I wrote adaptive attacks that got past them anyway, and wrote up why. |
| subagent-watch | VS Code extension that shows AI subagents while they run. Reads only documented local files, stores nothing sensitive, and the collector has a written security contract. |
| personal-dev-assistant | A terminal coding assistant with safety-checked file tools, sub-agent coordination and token budgeting. Dockerised. |
| multi-agent-hub | An agent that collaborates over a shared hub while treating every remote message as untrusted input — remote messages can never reach local tools. |
| tokeniser | VS Code extension that makes LLM token and context usage visible. SQLite indexing, atomic writes, permission-checked file handling. |
| deadline-bot | Discord bot that tracks course deadlines. Running in production on a cloud VM under systemd; 85 tests. Built for my class, and used by it. |
More in my repositories — including database design, ETL pipelines and neural networks built up from a single neuron to ResNet-50 transfer learning.
- I read the source when the docs are wrong. In the guardrails project I found that LiteLLM's documented
similarity_checksetting doesn't exist in the code and is silently ignored — which changed the whole design. - I write down what I couldn't make work. Every project has a limitations section. I'd rather state a weakness than have someone find it.
- I treat input as untrusted by default. Most of my projects have an explicit threat model, because most of them handle data that something else produced.
Python TypeScript SQL · PyTorch pandas FastAPI Docker SQLite PostgreSQL pytest · LLM agents, tool use, prompt-injection defence, evaluation suites
Teknikhögskolan Stockholm — AI and Machine Learning Engineering (HVE), HT25–VT27
Applied AI · Machine Learning · Python Frameworks · Databases · DevOps · Python Programming
Graded Väl Godkänd (Pass with Distinction) in Applied AI, with a special recognition for contributing course material and building the class's deadline bot.

