cv

Researcher

Independent Research — Agentic Misalignment & Alignment Faking

• Extended Lynch et al. (2025) agentic misalignment study across 22 models from 8B to frontier; found blackmail/leak/murder rates vary by model family and system-prompt framing, not model size, with DeepSeek highly vulnerable while latest OpenAI/Anthropic models stay near-zero (published on LessWrong)
• Replicated Koorndijk (2025), showing differential compliance in Llama 3 8B stems from lexical prompt sensitivity rather than genuine strategic reasoning (published on LessWrong)
• Attempted replication of Betley et al.'s emergent misalignment work; produced a negative result where 4-bit QLoRA fine-tuning caused code mode collapse rather than the reported persona shift

AI Safety Researcher (Part-time)

Jinesis AI Lab, University of Toronto

• Designed AF-Arena, a 5-axis evaluation arena for alignment faking in frontier LLMs; evaluated 9 models (GPT-5 series, Gemini 3, Claude 4.6, DeepSeek-R1, Grok 4.20) across 54,000 trials. Accepted as an oral at the ICML 2026 Workshop on Agents in the Wild; under review at NeurIPS 2026 (Evaluations & Datasets Track)
• Co-authored AM-Bench (submitted to NeurIPS 2026; with B. Schölkopf, Z. Jin et al.), a unified taxonomy and evaluation suite for agentic misalignment grounded in the fraud triangle (35 scenarios, 7 industries, 11 frontier models); built the modular evaluation harness, scenario set, and rationalization-scoring pass
• Conducted an 8,400-trial ablation study showing that elicitation conditions are axis-dependent: monitoring is necessary on both ablated axes, while explicit training criteria are additionally required on Axis E (sycophancy) but dispensable on Axis K (goal-guarding)
• Ran a 62,400-trial prompt-based mitigation sweep; identified and named three novel failure modes: behavioral masking, meta-alignment-faking, and self-aware alignment faking

Senior Software Engineer

Diedev GmbH

• Led 4-engineer team to deliver major product upgrade that increased company revenue by 35% within six months through improved feature delivery and system performance
• Designed and deployed production API gateway in Rust handling 1,000+ RPS with sub-100ms p99 latency, implementing intelligent routing, rate limiting, and distributed tracing
• Modernized critical Python 2 services to Rust, reducing build times by 60% and deployment frequency from weekly to daily releases
• Built scalable backend services and data ingestion pipelines, taking end-to-end ownership from design through production rollout and monitoring
• Improved CI/CD pipelines and observability infrastructure, enabling faster iteration cycles and smoother cross-team collaboration
• Developed full-stack features using TypeScript and Python, contributing across backend APIs, data pipelines, and frontend interfaces

Senior Software Engineer

Interswitch, Lagos, Nigeria

• Architected and built automated reconciliation system processing 100K+ daily transactions, reducing manual operations overhead by 25% and eliminating error-prone spreadsheet workflows
• Led technical scoping for enterprise clients alongside sales and product teams, translating business requirements into scalable solutions that closed 3 deals worth $200K+ combined
• Established engineering documentation standards and onboarding program, reducing new hire ramp-up time from 8 weeks to 4 weeks
• Built resilient, scalable features using Kafka/SQS messaging patterns for at-least-once delivery, ensuring data consistency across distributed services
• Implemented comprehensive logging, metrics, and tracing for operational excellence, maintaining high standards for documentation and enabling 40% faster incident resolution
• Owned full lifecycle of data infrastructure—from schema design and pipeline architecture through deployment, monitoring, and scaling to handle growing transaction volumes
• Worked cross-functionally with product, infrastructure, and data teams to design data models and ingestion strategies for new transaction types

Software Engineer

Minerva Technologies, Nigeria

• Designed and built internal admin platform with role-based access control and audit logging, enabling support teams to independently resolve 90% of merchant issues without engineering escalation
• Developed automated merchant transaction reversal system using event-driven architecture, increasing support throughput by 50%+ and reducing average resolution time from 2 hours to 15 minutes
• Shipped backend APIs and frontend features for fintech application serving 3M+ users, implementing payment processing, user authentication, and transaction history
• Built machine learning pipeline for digital identity verification and fraud detection using scikit-learn and custom feature engineering, reducing merchant verification steps and time by at least 2 hours per day
• Thrived in fast-paced startup environment, staying focused as priorities shifted and shipping scalable systems weekly with strong testing and code quality standards

Software Engineer

Adbasador Media Limited, Nigeria

• Built content management and delivery platform serving 50K+ daily active users, implementing caching strategies that reduced API response times by 40% and improved user engagement
• Led AWS cloud migration for monolithic application, implementing containerized microservices with CI/CD pipelines that reduced deployment downtime from 2 hours to under 10 minutes
• Developed real-time analytics dashboard and reporting tools using TypeScript/React and PostgreSQL, enabling business teams to track KPIs and make data-driven decisions
• Architected, designed, developed, and maintained 3 software solutions, re-engineering existing software quality to decrease page load times to less than 1 second
• Created optimized implementation plans ensuring smooth rollout resulting in 50% productivity enhancement; analyzed business processes and software requirements for implementation across multiple solutions

Vice President Education

TechyTalkers Toastmasters Club

• Organized meetings, assigned speech roles, and ensured members progress through their educational pathways
• Connected new members with mentors, tracked individual progress, and recognized member achievements
• Created the meeting schedules, planned contests, and ensured high-quality learning experiences

Volunteer

The Lagos Food Bank Initiative

• Facilitated outreaches and workshops on food distribution to indigent communities in Lagos, Nigeria
• Community outreach and social impact initiatives

Member

Data Science Club, CMU-Africa

• Led a team to participate in the Disability AI Hackathon where we used disability datasets to build solutions for disabled people in Africa
• Came out in the top 5 in the hackathon competition
• Applied machine learning and data science skills to address social challenges

M.Sc.

Carnegie Mellon University

Information Technology

• Machine Learning
• AI Research
• Computer Vision
• MLOps

B.Sc.

University of Nigeria

Biological and Data Science

• Data Science
• Biology
• Statistics
• Research Methods

Replication of Koorndijk (2025): Differential Compliance May Reflect Prompt Sensitivity, Not Strategic Reasoning

LessWrong

Differential compliance observed in Llama 3 8B is better explained by lexical prompt sensitivity than by genuine strategic reasoning.

Blackmail at 8 Billion Parameters: Agentic Misalignment in Sub-Frontier Models

LessWrong

Blackmail rates in sub-frontier open-weight models vary dramatically by model family rather than parameter count.

From 8B to Frontier: How System Prompts Control Whether AI Agents Blackmail, Leak, and Kill

LessWrong

Susceptibility to agentic misalignment across 22 models is driven by model family and system-prompt framing rather than model size.

AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment

Submitted to NeurIPS 2026

Zhang, E., Zhang, T.J., Ugwuanyi, C., Shi, J., Schölkopf, B., Jin, Z. — A taxonomy and benchmark for agentic misalignment grounded in the fraud triangle; 35 scenarios, 7 industries, 11 frontier models.

AF-Arena: An Evaluation Arena for Multi-Axis Alignment Faking in Frontier Language Models

Oral, ICML 2026 Workshop on Agents in the Wild; under review at NeurIPS 2026 (Evaluations & Datasets Track)

Ugwuanyi, C. (first author), Zhang, T.J., Schölkopf, B., Jin, Z. — Tests whether frontier LLMs fake alignment across five behavioral axes; characterizes three novel failure modes of prompt-based mitigations.

Hoppscotch API Live Sync

Production-ready feature for automatic API collection synchronization, enabling code-first API development workflow. Built intelligent framework detection for 7+ frameworks with real-time polling and persistent sync state.

• Open source contribution to Hoppscotch
• Intelligent framework detection (FastAPI, Express, Spring Boot, ASP.NET, Django, Flask, Rails, Laravel)
• Real-time polling with content hashing for change detection
• Service-oriented architecture with TypeScript
• Solved complex technical challenges including path parameter normalization and persistence
• Wrote comprehensive 6-part blog series documenting implementation

AI Village - Multi-Agent GitHub Issue Fixer

Real-time multi-agent system demonstrating LLM collaboration for fixing open-source GitHub issues. Implemented WebSocket streaming architecture with React frontend and FastAPI backend.

• Multi-LLM roundtable debate system with three distinct AI personas
• WebSocket streaming for instant UI updates during LLM inference
• Scout agent for GitHub issue discovery and analysis
• Human-in-the-loop review panel for quality control
• Evolved from Streamlit proof-of-concept to production-ready React + FastAPI system
• Integrated Ollama API for local and remote LLM inference

Evaluating Model Robustness to Adversarial Attacks

Research on CIFAR-10 and VGGFace2 that revealed ResNet-50 models drop from 87% to 9% accuracy under FGSM attacks, but adversarial training can recover 55% robustness.

• Adversarial robustness research on CIFAR-10 and VGGFace2 datasets
• Evaluated pretrained CNNs (ResNet-50, VGG16, MobileNetV2, EfficientNet)
• Implemented white-box and black-box attack scenarios
• Demonstrated adversarial training effectiveness
• Compared resilience across different model architectures

Multimodal Emotion Recognition

Designed and evaluated deep learning models combining audio, visual, and textual signals to improve emotion recognition robustness.

• Multimodal learning combining audio, video, and text signals
• End-to-end experimentation pipelines using Python and TensorFlow
• Rapid iteration on model architectures
• Analysis of trade-offs between complexity, accuracy, and real-world usability

MLOps Pipeline - End-to-End Data Engineering

Engineered comprehensive data engineering pipeline including data ingestion, orchestration, transformation, and analysis for production ML systems.

• Scalable data processing workflows with proper error handling
• Data validation, quality checks, and transformation pipelines
• Production-ready ML system infrastructure

AI-Assisted Coding Productivity Study

Conducted controlled self-experiment comparing AI-assisted and manual coding workflows, analyzing impacts on productivity, cognitive load, and learning outcomes.

• Controlled experimental design
• Productivity and cognitive load analysis
• Understanding developer interaction with AI tools

Comparative Analysis of Classification Models

Research to identify the best machine learning model for loan approval prediction based on accuracy and area under curve metrics.

• Evaluated Logistic Regression, Random Forest, SVM, KNN, MLP, and XGBoost
• Performance comparison using accuracy and AUC metrics
• Cross-validation and model selection

Plaid Shell Implementation

Implemented a fully featured Linux shell from scratch using C programming language, including command parsing, process management, job control, and I/O redirection.

• Systems programming in C
• Command parsing and execution
• Process management and job control
• I/O redirection and piping
• Understanding of Unix internals