Yasin Hessnawi · NorwayOpen to opportunities from summer 2027

Drag to look around · scroll to come closer · click anything that glows

Welcome to Yasin's lab

Agents at work. Everything passes the gate.

Initialising weights…

Yasin Hessnawi · Norway
Yasin HessnawiOpen to opportunities from summer 2027

Welcome to Yasin's lab

I build intelligent things, and the gates that keep them honest.

Software engineer and AI master's student at the University of Agder, graduating June 2027. I build AI agents and the checks around them, co-founded Safe Media AI, where I lead Declassifai, and right now I build LLM agent features at Green Norway.

755 bypass attempts · 0 protected writes

01

Projects

Safe agents3Swipe →

2026–now · UiA specialisation project · SFI CELECT Work Package 4

physgate

An agent harness for hardware–software co-design.

Read more

A deterministic orchestrator dispatches domain agents with paired reviewers. The core contribution is a physics gate of seven deterministic checks that stops physically invalid designs. Rules are enforced by hooks and an OS sandbox, not by prompts: in 755 bypass attempts, none wrote to a protected path. The learned escalation backends were tested in pre-registered experiments and rejected.

Results so far · Oct 2026

  • Physics gate on a sealed test set: caught 20 of 40 planted errors and failed none of 40 correct designs.
  • AI reviewers after calibration: accept 29 of 40 correct designs, up from 0 in the first test.
  • Reported as failed: one general reviewer beat the paired specialists (34 of 40 errors caught, 38 of 40 correct designs accepted, against 29 and 29).
  • Crash recovery: 10 of 10 clean restarts in each of 3 conditions.
  • Python
  • pint
  • SymPy
  • Claude Code
2025–now · Open core · MIT engine

Open Personas

An open-source agentic AI platform.

Read more

AI assistants with typed, versioned memory, a user-scoped knowledge graph, autonomy behind consent gates, real-time voice, MCP tool use across 300+ servers, ten model providers, and a turn-time safety gate.

What ships today

  • Four typed, versioned memory stores per persona: identity, self facts, worldview and episodic. Nothing is overwritten, and every write is audited.
  • One knowledge graph per user, shared by all personas, drawn as a memory map you can correct or forget from.
  • Schedules, initiative and event triggers, each behind approvals and spend caps.
  • Realtime voice over LiveKit that shares one memory with chat.
  • Telegram, Discord, Slack and email connectors, and 60 hand-written starter personas.
  • Python
  • FastAPI
  • Next.js
  • pgvector
  • LiveKit
  • MCP
2025–now

GLive

GitHub to Live.

Read more

Clone and run any GitHub project with one command: AI-powered setup, dependency management, security scanning and a PWA dashboard.

  • Go
  • Next.js
  • Docker

Ventures2Swipe →

2025–nowFeatured

Declassifai

An AI platform that finds and redacts personal data in documents and images, with 91%+ detection accuracy.

Read more

It uses end-to-end encrypted processing with no data storage, was piloted with Mattilsynet (the Norwegian Food Safety Authority), and runs on Azure Norway East. Awards and funding: NTNU Eureka Prize 2025 and an Innovasjon Norge commercialisation grant 2026.

  • AI/ML
  • React
  • Python
  • Azure
2025–nowFeatured

Safe Media AI AS

Co-founded the company behind Declassifai and further secure-processing tools.

Read more

I have been co-founder and CEO since January 2025.

  • Company
  • Secure processing

Research & learning systems6Swipe →

2026 · IKT460

RL agent for Chinese Checkers

2nd of 24 teams in the course tournament, with the highest average pin completion (9.5 of 10).

Read more

A residual network warm-started on heuristic and search-labelled positions, improved by self-play, playing through Gumbel Monte Carlo tree search.

  • PyTorch
  • RL
  • Gumbel MCTS
  • Self-play
2025–now

Graph Tsetlin Machine · Hex

Interpretable, clause-based winner prediction for Hex, CUDA-accelerated.

  • Tsetlin machines
  • CUDA
  • Hex
2026 · Independent research

Persona-RAG

Persona-vector extraction replicated on Gemma-2-9B, then shown not to transfer to multi-turn drift detection.

Read more

Typed-memory retrieval was evaluated on a 360-transcript probe suite.

  • Gemma-2-9B
  • RAG
  • Interpretability
MM-IMDb

Multimodal Genre Classification

Text and poster fusion reaching 59.8% F1-macro.

  • PyTorch
  • Multimodal
  • MM-IMDb

Deeper Neural Networks

Transformers, CNNs, mixture-of-experts and small agentic research systems.

  • PyTorch
  • Transformers
  • MoE

UiA RAG Chatbot

Hybrid BM25 plus dense retrieval with reranking.

  • RAG
  • BM25
  • Reranking

Smaller builds

  • STrackAI subscription tracker
  • CyberSight DWThreat-intelligence data warehouse with Kafka, PostgreSQL and Neo4j
  • 1.2.WatchAndroid
  • Youvies

02

About me

@ UiA
M.Sc. AI
years building
4+
university courses
21
projects shipped
12+
company co-founded
1

Now

  • AI Developer Intern08/2026–now
    Green Norway, Kristiansand

    Extending Lumi, the LLM agent in the GreenSight energy platform, so it fetches and processes user data through tools without the model ever seeing the values. Building AI automation of the development workflow in GitHub Actions: code review, fixing blocking issues, resolving merge conflicts, auto-merge. Migrating integrations to Azure Functions with 500+ automated tests.

  • Co-founder and CEO01/2025–now
    Safe Media AI AS (Declassifai)

Before

  • Teaching Assistant and Student Assistant2024
    NTNU Gjøvik

    With the LAOS teaching assistant certificate.

Education

  • M.Sc. in Technology (Sivilingeniør), Artificial Intelligence2025–2027
    University of Agder
  • B.Eng. Data Engineering, Programming and Cybersecurity2022–2025
    NTNU

    Grade A thesis and the Eureka Prize.

Languages

  • Arabic (native)
  • Norwegian (C1)
  • English (C1)

Toolkit

Languages
  • Python
  • Go
  • TypeScript/JavaScript
  • Java
  • Kotlin
  • C/C++
  • SQL
AI/ML
  • PyTorch
  • NLP
  • Computer vision
  • RL
  • Tsetlin machines
  • RAG
  • LLM agents
  • MCP
Web
  • React
  • Next.js
  • FastAPI
  • Node
Cloud
  • Azure
  • AWS
  • Docker
  • Kubernetes
  • Terraform
  • GitHub Actions
Data
  • PostgreSQL
  • pgvector
  • MongoDB
  • Neo4j
  • Kafka

Spec-driven development with Claude Code.

Download résumé

03

Notes from the coursework

Short write-ups of what I learned in 21 university courses, from the first programming course to Tsetlin machines.

04

Let's build something.

Questions, collaborations, PhD or job opportunities: write to me and I'll reply when I can.

Designed and built in Norway · © 2026