Languages
PythonCJavaScriptTypeScriptJavaSQLJuliaR
and hello.
In Amharic, ሰላም (sälam) means peace. It's how you greet someone in Ethiopia and Eritrea.I'm Kidus Dereje Zewde, a Computing Science and Economics student at the University of Alberta, graduating June 2026, and a Founding Engineer at Scam AI. My work sits where research meets production: nine papers on deepfake and document forensics, and the detection systems that put them to use.
I care about the whole stack, from model architecture to the interface people touch, and I'm drawn to problems where rigorous engineering and creative thinking both matter. Off the clock: film photography and good coffee.
Experience
-
Founding Engineer
- Built core detection pipelines for Eva V1.6, Scam AI's multi-modal deepfake and forgery engine: face-swap, lip-sync, GAN fingerprinting, diffusion signatures, document forgery localization, voice clones across image, video, document, and audio. SOC 2 Type II, GDPR-compliant; catches 98.2% of deepfakes at sub-4-second inference.
- Shipped Halo with Qualcomm: the first on-device, real-time deepfake detector for live video calls (Zoom, Teams, Meet) running ~4 checks/sec with zero video upload.
- Hands-on technical point of contact for enterprise clients integrating the REST detection API into onboarding, claims, and content-moderation workflows.
- Authored 9 peer-reviewed and arXiv papers on synthetic-media forensics that directly informed Eva's model selection and training-data strategy.
-
Machine Learning Engineer
- Built voice-clone and synthetic-audio detection models (cross-language, ElevenLabs/PlayHT/Azure TTS, splice/pitch/speed manipulations) reaching 98.5% accuracy in under 3 seconds per clip, served through real-time and batch REST endpoints.
- Worked across the CheckReality.ai forensic stack: GAN fingerprints, diffusion signatures, frequency-domain anomalies, metadata forensics, C2PA validation, document forgery, liveness.
- Engineered a synthetic scam-data pipeline (LangChain, ElevenLabs, Qwen-MT) producing training samples in 14 languages.
- Designed a multi-agent scam-call scoring system (Deepgram, LiveKit, FastAPI, fine-tuned OpenAI model) at 80% success rate, and an agentic SMS scam-detection API on a fine-tuned Qwen model.
- Implemented CAM explainability heatmaps (PyTorch, EfficientNet) for audit-ready forensic reports.
-
Machine Learning Intern
- Fine-tuned YOLOv5 on a car-theft image dataset, improving accuracy by 20%.
- Engineered ETL pipelines for ML data ingestion.
- Automated data validation and augmentation for continuous training.
Publications
Nine papers on synthetic-media forensics, 2025 to today.
- Analyzed a complete corpus of 10,211 inbound scam and spam calls (913 hours of audio, 330,956 transcribed turns from 5,780 distinct numbers) collected over 54 days by an AI voice-agent honeypot.
- Ran a randomized experiment across 1,823 calls using ten fictitious caller identities, finding scammers spent more conversational turns working older-sounding targets.
- Showed that scam escalation is predictable early from a caller’s opening lines alone, benchmarking classifiers up to 0.87 ROC-AUC by the eighth line on a caller-disjoint split.
audiobehavioraldataset
- Introduced CallScreenBench, a benchmark for on-device small language models (0.6B to 4B, 4-bit) that answer unknown inbound calls on their owner’s behalf, where the caller holds the goal and may be an adversary.
- Defined five automated call-and-note measure groups motivated by owner endorsement, each paired with a counter-metric and an uncertainty estimate, plus three guardedness diagnostics for a toolless proxy holding no credentials.
- Showed that larger checkpoints score higher on service, recall, and plausibility while triage discrimination follows a different ordering, and that scripted degenerate agents expose floors a single pass/fail score would hide.
llm-evalbenchmarkbehavioral
- Curated 10,217 confirmed GPT-image-2 images from Twitter/X within six days of release using a multilingual (English, Japanese, Chinese) text-heuristic pipeline and browser-automated badge verification.
- Designed a multi-stage data curation pipeline labeling over 27,000 raw image records with high-confidence provenance.
- Documented that C2PA cryptographic content credentials are systematically stripped by Twitter’s CDN on upload, with broad implications for AI image provenance research.
dataset
- Demonstrated that GPT-Image-2 inpainting erases the visual boundary between authentic and AI-edited document images, enabling seamless single-field manipulation at near-zero cost.
- Benchmarked three forensic judges (TruFor, DocTamper, GPT-Image-2 self-judge) on a newly released dataset of AI-inpainted receipts and financial documents; forensic AUC collapsed from 0.962/0.852 to 0.599/0.585 on GPT-Image-2 inpainting.
- Released the full dataset, curation pipeline, and four-judge evaluation protocol to support future AI forensics research.
documentsbenchmark
- Introduced an open-source infrastructure and dataset (final_dataset_v1, 12 hours) mapping real human iris trajectories from webcams and replaying them on a 3D simulator via headless browser automation.
- Applied the pipeline to script-reading detection during video interviews, showing synthetic trajectories preserve the temporal dynamics of real eye movements (KS D < 0.14).
- Characterized a visual sensitivity limitation at reading-scale movements caused by the absence of coupled head motion, establishing design baselines for future behavioral simulators.
datasetbehavioral
- Led the first large-scale zero-shot benchmark of AI-generated image detectors: 23 pretrained models across 12 datasets and 2.6 million image samples.
- Identified critical generalization gaps — detector performance is highly context-dependent (Spearman ρ as low as 0.01) and training data alignment outweighs architecture.
- Developed deployment guidelines showing modern generators (Midjourney, Flux) frequently defeat existing detectors, with a framework for threat-specific model selection.
benchmark
- Benchmarked GPT-4o, Gemini, and Llama 3.2 for detecting document fraud across diverse forgery types.
- Demonstrated that top-performing LLMs show superior zero-shot generalization over traditional SVM and CNN baselines for out-of-distribution forgeries.
- Found that model size and advanced reasoning show limited correlation with detection accuracy, while providing a foundation for interpretable, scalable fraud mitigation.
documentsllm-eval
- Benchmarked 12 state-of-the-art multi-modal LLMs (including GPT-4o, Gemini 2, Claude 3.7) for zero-shot deepfake detection across multiple datasets.
- Conducted ablation studies investigating the impact of model size, version updates, and reasoning capabilities on detection performance.
- Analyzed failure modes and interpretability through score distribution analysis and reasoning pathway examination.
deepfakellm-eval
- Investigated the vulnerability of deepfake detection methods to real-world data manipulations, particularly super-resolution post-processing.
- Contributed to the creation of a novel real-world faceswap dataset to benchmark deepfake detectors in practical settings.
deepfake
Skills
Frameworks & tools
Next.jsReactDjangoFastAPIFlaskTailwind CSSAWS (S3, Lambda, Bedrock, DynamoDB, CloudFront)DockerKubernetesGit
ML & data
PyTorchTensorFlowScikit-learnPandasNumPySeabornPower BI
Databases
PostgreSQLMongoDBCassandraPrisma ORM
Dev & testing
JenkinsRESTful API designAgile methodologiesData visualization (R, Python)JestPytestPostmanSelenium
Certifications
- UAlberta Certificate of Innovation and Entrepreneurship
- Google Cybersecurity Professional
- Google IT Automation with Python
- UAlberta Reinforcement Learning Specialization
Education
BSc Computing Science, Minor in Economics
University of Alberta
June 2026