Skip to content
View Lullow's full-sized avatar

Block or report Lullow

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Lullow/README.md

Elia Cherrou

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.

🎯 Available for LIA (internship)

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.


Currently building

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.


What I've built

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.


How I work

  • I read the source when the docs are wrong. In the guardrails project I found that LiteLLM's documented similarity_check setting 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.

Tech

Python TypeScript SQL · PyTorch pandas FastAPI Docker SQLite PostgreSQL pytest · LLM agents, tool use, prompt-injection defence, evaluation suites

Education

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.

Contact

LinkedIn · elias-cherrou@live.se

Pinned Loading

  1. privacy-request-manager privacy-request-manager Public

    GDPR tool that drafts, sends and tracks personal-data removal requests to Swedish data brokers - React SPA on an async FastAPI backend.

    JavaScript

  2. llm-guardrails llm-guardrails Public

    Three layers of prompt-injection defence in front of local LLMs via LiteLLM, with a 22-attack evaluation suite.

    Python

  3. multi-agent-hub multi-agent-hub Public

    Agent that collaborates over a shared HTTP hub while treating every remote message as untrusted input, isolated from local tools.

    Python

  4. personal-dev-assistant personal-dev-assistant Public

    Terminal-based AI coding assistant with safety-checked file tools, sub-agent coordination and token budgeting.

    Python

  5. personal-life-agent personal-life-agent Public

    Conversational terminal assistant for tasks, calendar events and reminders - understands free-text Swedish and English.

    Python

  6. tokeniser tokeniser Public

    VS Code extension that makes Claude Code token and context usage visible, from documented local sources only.

    TypeScript