About

Building at the intersection of machine learning and software engineering.

ሰላም and Hello! I'm Kidus Dereje Zewde — a Computing Science + Economics student at University of Alberta (graduating June 2026), currently working as a Founding Engineer at Scam AI. My work sits at the boundary between research and production: I've published 4 papers on deepfake and AI-generated content detection, and I build systems that put those ideas into practice.

I care about the full stack — from model architecture to user-facing product — and I'm drawn to problems where rigorous engineering and creative thinking both matter.

Education

BSc Computing Science + Economics Minor with additional Certificate in Innovation and Entrepreneurship University of Alberta

Expected June 2026

Experience

Contributed to the development of Scam AI's Eva-v1 deepfake and AI-generated media detection models, building ML pipelines to detect face swaps, expression and facial-attribute manipulations, and synthetic imagery from Stable Diffusion, DALL·E, Midjourney, and Flux across image and video, reaching 98.2% accuracy with confidence scores and manipulated-region heatmaps served via a RESTful API at sub-4-second inference for KYC and content-moderation use cases
Worked across CheckReality.ai's enterprise forensic detection stack including AI-generated image detection (GAN fingerprints, diffusion signatures, pixel-level noise, frequency domain anomalies, metadata forensics, C2PA credential validation), document forgery analysis (bank statements, pay stubs, IDs), active liveness and age estimation for identity verification, and remote interview integrity — and implemented CAM-based explainability outputs using PyTorch and EfficientNet to support SOC 2 Type II compliant, audit-ready forensic reporting
Built Scam AI's voice clone and synthetic audio detection models, identifying cloned voices across languages and accents, text-to-speech from ElevenLabs, PlayHT, and Azure TTS, and audio manipulations such as splicing, pitch, and speed alterations — reaching 98.5% accuracy in under 3 seconds per clip, with real-time and batch REST API endpoints powering live call verification for vishing prevention

Engineered a synthetic data generation pipeline using LangChain, ElevenLabs, and Qwen-MT to produce high-quality scam samples in 14 languages for ML model training.
Designed a multi-agent AI system using Deepgram, LiveKit, FastAPI, and a fine-tuned OpenAI 4.1 model to transcribe and score potential scam calls, achieving 80% success rate.
Developed an agentic SMS scam detection API using FastAPI and a fine-tuned Qwen3.2-32B model via LangChain for adaptive real-time detection.
Implemented CAM visualization for a deepfake detection model using PyTorch and EfficientNet to produce interpretable AI tampering heatmaps.

Fine-tuned a pre-trained YOLOv5 object detection model on a specialized car theft dataset, increasing accuracy by 20%.
Engineered ETL pipelines for ML data ingestion, streamlining feature processing for continuous model training and evaluation.
Implemented automated data validation and augmentation scripts to ensure high-quality, consistent data streams.

Publications

Skills

Languages

Python
TypeScript / JavaScript
Java
C / C++
SQL
Swift
R

Frameworks

React / Next.js
SvelteKit
Django / FastAPI
PyTorch
TensorFlow
scikit-learn

ML / Data

NumPy / Pandas
HuggingFace
OpenCV
Gemini API
RAG Pipelines
Matplotlib

Databases

PostgreSQL
Firebase
MongoDB
MySQL
Prisma
Supabase

DevOps

Docker
AWS (Lambda, S3, Bedrock)
GitHub Actions
Vercel
Git