Open to AI/ML roles · available full time nowBased in Linz, Austria

Ahmed
Mohammed

AI SystemsLLM ProductsAgent Tooling

AI/ML engineer. I build ML systems I can stand behind — verified error bars in research, eval harnesses in production.

View work Download CV

orSchedule meetingGet in touch

Years in AI/ML

99.03%

OOD Detection AUROC

+6.5pp

Gain over baseline (stable)

99.03% AUROC = CIFAR-10 thesis (binary, single-class airplane-vs-rest, 3-seed mean) · 0.8673 AUROC = industrial FTI_Zer0P 5-fold baseline · +6.5pp = gain from separation loss

Thesis · live dataseed 42

Diffusion anomaly scores separate ID from OOD

τ @ TPR 95

anomaly score →

In-distribution · airplane
Out-of-distribution · 9 other classes

Headline figure · mean of 3 seedsAUROC 99.03% ± 0.07

This run · seed 42

AUROC: 98.98%
FPR@95: 4.7%
AUPRC: 99.87%

CIFAR-10, seed 42, λ = 0.02, K = 100 · 1,000 ID vs 9,000 OOD test images

Explore the interactive lab

JKU Linz, Austria

Supervised by Prof. Sepp Hochreiter

PROFACTOR / FTI_Zer0P

99.03% AUROC thesis result (binary, single-class)

AI QC platform shipped to production

// 01 - About

AI/ML Engineer - research that ships.

Transitioned from mechatronics engineering to AI/ML - bringing a hardware and systems perspective to computer vision and production ML.

I am an AI/ML engineer who holds an M.Sc. in Artificial Intelligence from JKU Linz (graduated September 2026; Defensio 25 Sep 2026), supervised by Prof. Sepp Hochreiter. My thesis was graded 1, the top mark.

My work spans the full stack: from class-conditional separation loss for diffusion-based OOD detection to industrial computer vision pipelines evaluated under rigorous cross-validation. I also ship LLM products and the deterministic tooling that keeps AI coding agents honest. 99.03% +/- 0.07% AUROC (binary airplane-vs-rest, 3-seed mean) · 0.8673 AUROC (FTI_Zer0P 5-fold).

I also build and ship — most recently Sihem, an LLM personal-mentor side project accepted into the tech2b incubator, and before that Faultrix, an AI quality-control platform I took from zero to production on my own. Both taught me that production reliability is its own kind of rigor.

Location

Linz, Austria

Supervisor

Prof. Sepp Hochreiter

Current work

Sihem — LLM side project (tech2b incubator)

Open to

AI/ML Roles - Research Collaborations

Education

M.Sc. in Artificial Intelligence

Oct 2020 – Sep 2026 (graduated 25 Sep 2026)

Johannes Kepler University Linz

Thesis (graded 1): conditional diffusion models for OOD detection - 99.03% +/- 0.07% average AUROC (binary airplane-vs-rest, 3-seed mean).

B.Sc. in Mechatronics Engineering

Feb 2015 - Jan 2018

Eastern Mediterranean University, Famagusta, Cyprus

Thesis: SCARA robotic system for dynamic object tracking.

Skills & Stack

Deep Learning & CV

PyTorchPyTorch LightningDiffusion Models (DDPM, UNet)CNNsTransformersOOD DetectionYOLOv8Defect Inspection

GenAI & Agentic AI

LLM API IntegrationMulti-Provider Routing & FailoverPrompt EngineeringStructured Output Validationpgvector Retrieval

ML & Data Science

HydraAblation StudiesAUROC/FPR95Cross-ValidationMulti-Seed RunsRoboflow

Deployment & MLOps

PythonGitLinuxDockerREST APIsGitHub Actions CI/CDSupabaseTypeScriptReact

Agent Tooling

mergeproofClaude CodeGitHub CopilotDeterministic checks

Languages

Arabic

Native

English

B2 - Professional

German

B1 CEFR - preparing for B2

// 02 - Projects

Research that ships.

Current project · Live beta

Sihem — AI Mentor Assistant

A proactive personal-mentor assistant on Telegram that plans your week, sends morning briefings and evening reviews, and fires smart check-in triggers. Built on Supabase edge functions (Deno/TS) with a generic activity model (workouts, medication, study sessions, custom habits — no schema change per type), a deterministic generators→arbiter→composer decision engine for adaptive nudging, pgvector long-term memory with RAG, multi-provider LLM routing (GLM/Gemini/Groq) with provider failover, multimodal meal-photo vision, TTS voice reminders (Cloudflare MeloTTS), Google Calendar sync, optional Alexa Echo integration, invite-code gated beta access, and an ops dashboard with cost rollups. Now in beta — try it on Telegram at @sihem_ai_bot.

Deno / TSSupabase Edge Functionspgvector RAGLLM Routingpg_cronTelegram

Telegram bot · PWA · pgvector memory

Case study →Telegram →PWA →

Open source · Agent tooling

mergeproof

Decides whether a pull request written by an AI agent is safe to merge. Plain deterministic code that never calls an LLM, so the agent being judged cannot talk its way past it.

0 LLM calls

JavaScriptZero dependenciesLLM-free gatenpm

Code →

Master's Thesis - JKU Linz

OOD Detection Framework

Conditional diffusion models as generative classifiers achieving 99.03% +/- 0.07% average AUROC on a within-CIFAR airplane-vs-rest binary split (single ID class; not comparable to multi-class CIFAR-10 OOD benchmarks). Introduced class-conditional separation loss improving average performance by 6.5 percentage points with high seed stability.

99.03% +/- 0.07% Avg AUROC

Code →Case study →

Master's Thesis Track 2

InkjetOOD Defect Detection

YOLO+CDM quality classification pipeline evaluated on the public FTI_Zer0P benchmark under strict 5-fold cross-validation.

0.8673 AUROC (5-Fold CV)

Code →Case study →

Previous project · 2025-2026

Faultrix

AI-powered construction quality control SaaS. Analyzed building photos and generated ONORM-compliant reports in under 1 minute - SHA-256 evidence chain, DSGVO compliant, AES-256 encryption. Built solo from zero to production.

<1 min per report

Next.jsConvexOpenAIClerkCloudflare R2Stripe

faultrix.com →Case study →

Karunya Institute of Technology and Sciences · BCI Research

EEG Signal Classification

LSTM, Bi-LSTM, and GRU architectures for motor imagery classification in brain-computer interface applications.

Code →

More open source

Elbostan Pazar bot — a Messenger bot that answers a real shop's customers about prices, hours and delivery. · brainstorm2idea + idea2prototype — zero-dependency tools that take a fuzzy idea to a linted brief, then to a tested prototype, with any coding agent.

// 03 - Proof

Selected evidence.

Public artifacts, reports, and repositories behind the portfolio claims.

Proof

Master thesis PDF

99.03% +/- 0.07% AUROC

Open →

Proof

Industrial OOD baseline (thesis)

0.8673 +/- 0.0230 AUROC · FTI_Zer0P 5-fold

Open →

Case study

Early technical report (2024)

Negative result — motivated the thesis approach

Open →

Proof

DiffusionOOD repository

CIFAR-10 OOD implementation

Open →

Proof

InkjetOOD repository

Industrial transfer codebase

Open →

Proof

Hugging Face model weights

Public checkpoints and artifacts

Open →

Proof

Research poster

Visual summary of the thesis

Open →

Case study

DiffusionOOD case study

Problem, method, results, lessons

Open →

Case study

InkjetOOD case study

Industrial transfer, results, and lessons

Open →

Proof

Sihem — live beta

Running system — Telegram bot and PWA, open to try

Open →

Case study

Faultrix case study

Research-to-product build story

Open →

// 04 - Experience

Where I've worked.

May 2026 Present

Solo Developer & Project Lead

Sihem (side project) · tech2b incubator, Linz

Design and run the full cloud stack behind a habits and daily-routine assistant, delivered as a Telegram bot and an installable web app: managed PostgreSQL with row-level security, Deno/TypeScript serverless functions, scheduled jobs, pgvector memory, and GitHub Actions deploys. LLM calls route across several providers with escalation and failover.

Live beta - @sihem_ai_bot

Apr 2024 Jul 2026

M.Sc. Thesis Researcher, Machine Vision

JKU Institute for Machine Learning with PROFACTOR GmbH · Linz / Steyr

Conditional diffusion models as generative classifiers for OOD detection under Prof. Sepp Hochreiter. A class-conditional separation loss cut seed-to-seed std from 11.07 to 0.07 points while raising CIFAR-10 AUROC from 92.52% to 99.03%. Applied in project Zer0P (zero-defect inkjet printing on building components) as a two-stage YOLOv8 + conditional-diffusion pipeline, with an eight-method 5-fold comparison.

Thesis graded 1 (top mark)

Aug 2023 Oct 2023

AI Research Intern

Karunya Institute of Technology and Sciences · Coimbatore, India

RNN/CNN architectures for EEG motor imagery classification with hyperparameter optimization.

Sep 2018 Oct 2020

Electronics Technician

Ledzone · Karaköy, Istanbul, Turkey

PCB assembly and soldering, LED and motor-driver circuit build-up, reading schematics and layouts, and functional test, fault finding and final inspection with multimeter, oscilloscope and function generator.

// 05 - Research

Reports & Thesis

Problem

How can a model recognize when an image does not belong, instead of acting overconfident?

Contribution

A separation loss that forces the model s competing class explanations to move apart.

Result

99.03% +/- 0.07% AUROC on CIFAR-10 airplane-vs-rest (binary, single ID class, 3-seed mean), with a +6.5pp gain over the non-separated baseline.

Why it matters

The score became not only higher, but dramatically more stable across seeds and easier to trust.

Master's Thesis (Track 1) - JKU Linz - 2026

Conditional Diffusion Models as Generative Classifiers for Out-of-Distribution Detection in Inkjet Print Quality Control

A diffusion model compares how well different class explanations reconstruct an image. The new separation loss pushes those explanations apart, making the anomaly score clearer and more stable. 99.03% +/- 0.07% AUROC, +6.5pp gain, baseline 92.52% +/- 11.07%, external OOD 90.50%-96.97%.

Supervisor: Prof. Sepp Hochreiter - Assistant: Claus Hofmann, MSc - JKU Linz - 2026

99.03%

AUROC - CIFAR-10 Average

+/-0.07% variance (3 seeds)

Separation-weight sweep · Threshold calibration · Reconstruction error

Separation weight sweep: best stability at lambda=0.02

Score distributions at calibrated operating threshold

Reconstruction error gap peaks at intermediate noise

Novel contribution - separation loss

Plain idea: train the model to make its two explanations visibly different. The gap becomes the anomaly signal.

1Learn to denoiseReconstruct images under each class condition.

2Push explanations apartIncrease the distance between conditional predictions.

3Score the gapUse the reconstruction-error difference as the OOD signal.

Total loss=denoising loss+lambda * separation loss

Latest thesis result: lambda = 0.02 reached 99.03% +/- 0.07% AUROC, a +6.5pp gain over no separation loss.

Master's Thesis (Track 2) - JKU Linz - 2026

Rigorous Cross-Domain Transfer & FTI_Zer0P Benchmark

The public YOLO+CDM print-quality pipeline was evaluated on FTI_Zer0P with strict 5-fold cross-validation, showing where separation loss transfers and where small manufacturing textures behave differently. 0.8673 +/- 0.0230 AUROC.

Public Benchmark - InkjetOOD Codebase - FTI_Zer0P Public Dataset

0.8673

AUROC - 5-Fold CV Baseline

Rigorous Public Evaluation

Dataset Overview: Industrial thermographic frames · Zero-defect samples · Defect categories · 5-Fold Cross-Validation