Chidubem Onwuchuluba | Backend & Applied AI Engineer

01 / About

Turning ideas into
working systems.

Engineer, researcher and teaching assistant at Georgia State.

Usually found shipping something

Backend and applied AI engineer. I build the agents, the APIs, and the infrastructure underneath.

I'm an M.S. Computer Science student at Georgia State University on the thesis track, concentrating in Machine Learning and Software Engineering. Most of what I build sits behind the interface: multi-agent systems that talk to each other over A2A and reach their tools through MCP, a Django and PostgreSQL platform on AWS, and payment APIs written in C# and Python.

On the research side, I work on adversarial machine learning, attacking and defending the face authentication pipelines used in digital KYC, and I've co-authored two papers on adversarial defenses.

I also teach: I support 110+ undergraduates in Computer Organization Programming and run thesis-formatting workshops for graduate students, for which I built an automated format-compliance checker.

3.92

Graduate GPA

−45%

API response time at Paystack

4

AI agent systems built

0.997

AUROC on adversarial inputs

Right now

Building agentic systems with the OpenAI Agents SDK, A2A and MCP, plus tools like CodeLens for assessing how engineers work with AI.
Researching adversarial defenses for face authentication pipelines.
Teaching 110+ students at GSU and running thesis-formatting workshops for graduate students.
Studying for my M.S. in Computer Science (3.92 GPA, thesis track).

Atlanta, Georgia

--:--EST

My local time

02 / Experience

Where I've worked.

From payments infrastructure to the classroom: building things people rely on.

GSU
Georgia State University
Graduate Assistant
Aug 2025 – Present
- Teach and support 110+ undergraduates in Computer Organization Programming: running labs, debugging student code and configuring Linux/Windows lab environments.
- Wrote Python automation scripts that streamlined grading and lab administration, freeing time for one-on-one student support.
- Design and lead thesis and dissertation formatting workshops for graduate students across the College of Arts and Sciences, translating APA and ASA template requirements into plain, step-by-step guidance.
- Built and shipped a thesis/dissertation format-compliance checker: a deterministic two-tier design (XML-structural checks plus PDF-rendered geometry) backed by a manifest of 48 verified evidence checks, with mandatory LibreOffice rendering so every finding is visually verified, and wrapped it in a FastAPI app so students can self-check before submission.
- 110+ students
- 95% satisfaction
- −70%+ review time
- 94% checker accuracy
- 48 evidence checks
- Python
- FastAPI
- LibreOffice
- XML / PDF
- Linux
- Teaching
PS
Paystack · YC W16, a Stripe company
Software Engineer
May 2024 – Aug 2025
- Built and optimized RESTful APIs in a platform processing 2M+ daily transactions at 99.9% uptime; implemented a Redis caching layer that cut response time by 45%.
- Contributed to a multi-tenant SaaS architecture serving 50,000+ merchants, building containerized microservices on Docker and Kubernetes.
- Improved CI/CD pipelines on Jenkins and GitLab CI, raising deployment frequency by 60% with zero-downtime blue-green releases.
- Profiled slow queries across high-throughput payment flows and reduced p99 latency through index tuning and query restructuring.
- Worked in C# alongside Python across a mixed AWS and Azure footprint, collaborating with PMs and cross-functional teams from requirements through production.
- 2M+ daily transactions
- 99.9% uptime
- −45% response time
- +60% deploy frequency
- 50K+ merchants
- Python
- C#
- Redis
- Docker
- Kubernetes
- AWS
- Azure
- Jenkins
- GitLab CI
RE
Rotawn Energy
Software Engineering Intern
Jan 2024 – May 2024
- Built data visualization dashboards for requisition and expense tracking, enabling real-time reporting that contributed to a 3% cost reduction in 6 months.
- Developed and deployed a Visitor Management web app with QR-based authentication and SMS notifications, cutting manual processing time by 40%.
- Maintained 99.5% uptime through proactive monitoring with AWS CloudWatch, database query optimization and automated backup pipelines.
- −3% costs in 6 months
- −40% manual processing
- 99.5% uptime
- Python
- Django
- JavaScript
- Chart.js
- AWS CloudWatch
Self-directed
Independent Study & Certifications
May 2023 – Dec 2023
- After graduating, focused on building new skills full time, working through courses in machine learning and full-stack web development and earning certifications.
- Machine learning
- Python
- Django
- Full-stack development

03 / Projects

Things I've built.

Agent systems, a production platform and research.

01 AI Agents · Multi-Agent Systems

Atrium: Multi-Agent Trip Planning Network

A network of specialist LLM agents that plan a trip together. Each agent runs as an independent FastAPI service; they discover each other and delegate work over the Agent2Agent (A2A) protocol, while each reaches its own tools over MCP.

Exposed external tools and data through MCP servers, keeping tool access separate from agent reasoning.
Built every agent on the OpenAI Agents SDK, but kept the protocol layer framework-agnostic, so an agent built on a different SDK could join the network over A2A without changes elsewhere.
Async Python end to end, so agent-to-agent and tool calls run concurrently.

3 agentsConcierge, Travel, Local Experience

A2A + MCPVendor-neutral protocol layer

Python asyncio
FastAPI
OpenAI Agents SDK
A2A
MCP
Pydantic
mypy strict

View code Read the write-up

codelens · candidate session

CodeLens candidate IDE with file tree, editor, test output and an AI agent chat

reviewer · agentic coding score

CodeLens reviewer page showing an Agentic Coding Score of 64 out of 100 with pillar breakdown

02 Agentic Coding · Full Stack

CodeLens: Agentic Coding Assessment Platform

A browser-based assessment platform that scores how engineers investigate, prompt, verify and decide when to trust AI agents, with assessor, candidate and reviewer flows, all inside one fully instrumented sandbox.

Provisions an isolated E2B sandbox VM per session in about 5 seconds: clones the repo, installs dependencies, pre-authenticates Claude Code and Codex, and allocates a real PTY.
Candidate IDE with file tree, Monaco editor, xterm.js terminal and an in-app tool-calling agent chat streamed over NDJSON.
Instruments every action into Supabase Postgres (edits, terminal commands, test runs, prompts with token counts, and Claude Code tool calls via hooks), with secrets redacted before persistence.
A deterministic Agentic Coding Score across four weighted pillars, where every signal traces to logged events and no LLM judges the candidate.

~5sSandbox VM per session

4 pillarsOutcome · AI · Comprehension · Process

Next.js 16
TypeScript
E2B
Supabase
Monaco
xterm.js
Claude Code
Codex

View code

03 Research · Adversarial ML · Security

Defending End-to-End Face Authentication Pipelines

Digital KYC onboarding at banks, fintechs and payment platforms gates identity with a liveness model chained to a face recognition model, and an attacker can fool both in one shot. I built that pipeline, attacked it, and proposed a defense that reads the models’ own internal activations.

End-to-end pipeline: a ResNet-18 face anti-spoofing model gating a FaceNet face recognition model.
Pipeline-aware attacks against both stages: FGSM, PGD, Carlini-Wagner, and a joint multi-objective PGD that defeats liveness and identity matching at the same time.
Internal-activation fingerprinting: a 1,411-dimensional feature vector from intermediate layers flags adversarial inputs before an authentication decision is made.
Tested cross-paradigm transferability against DeePixBiS, a depth-based liveness model; extends earlier work on an adaptive multi-stage detector (85–93% detection at 6.4 average forward passes).

0.997AUROC detecting adversarial inputs

1,411-dActivation fingerprint

PyTorch
ResNet-18
FaceNet
DeePixBiS
FGSM · PGD · C&W

offplan.cwlagos.com

CW Lagos Off-Plan homepage: Lagos off-plan developments, intelligently tracked and clearly priced

CW Lagos Off-Plan on mobile: curated developments with the Metropolitan Towers listing

04 Full Stack · Cloud · Data Platform

CW Lagos Real Estate Operations Platform

A full-stack cloud platform I architected and still run for CW Lagos, tracking 300+ off-plan properties for 700+ monthly visitors on AWS (EC2, S3, CloudFront, ELB, CloudWatch) with auto-scaling, a CDN and monitoring.

Optimized PostgreSQL queries and added connection pooling and load balancing, reaching 98% uptime and 2-second page loads while cutting infrastructure cost by 25%.
Diagnosed recurring OOM kills on a t3.micro through memory profiling and right-sized to t3.medium, ending production crashes.
Built a Celery-driven Airtable sync pipeline; fixed a crash on blank fields, eliminated duplicate sync scheduling and made stale-record pruning safe.
Added Pinecone vector search for semantic property lookup (an earlier version ran RAG over listings with ChromaDB); CI/CD with GitHub Actions and n8n workflows.

300+Off-plan properties

700+Monthly visitors

98%Uptime · ~2s page loads

−25%Infrastructure cost

Django
PostgreSQL
Celery
Airtable
Pinecone
Next.js
AWS
GitHub Actions
n8n

Visit live site

DDispatch AIbot · online

Schedule a meeting with John tomorrow at 3pm

→ CalendarAgent

Done ✓ “Meeting with John” is on your calendar for tomorrow, 3:00 PM.

Remind me to call Mom in 1 hour

→ ReminderAgent

Got it. I'll remind you in 1 hour.

Message

05 AI Agents · Automation

Dispatch AI: Telegram Personal Assistant

A dispatcher-plus-specialists assistant that lives in Telegram. A routing agent reads each message and hands off to one of seven specialists: Gmail, Calendar, Tasks, web search, reminders, notes and macOS control.

Gmail agent with correct MIME threading so replies land in the right conversation; Calendar and Tasks agents with full CRUD and ISO 8601 handling for reliable dates and time zones.
Google OAuth 2.0 with automatic token refresh, and live web search through the Brave Search API.
Stateful per-user conversation memory with 30-minute session expiry, so multi-turn requests keep their context.
Async push reminders with APScheduler so the bot messages you proactively, plus a macOS control agent that runs local system actions from chat.

7Specialist agents

30 minConversation memory

Python
OpenAI Agents SDK
Telegram Bot API
Google OAuth 2.0
Brave Search API
APScheduler

View code

example output · zsh

❯ python main.py

=== Autonomous Code Review System ===

Enter choice (1 or 2): 1

Enter file or directory path: src

Running all 4 reviewers in parallel...

✓ STATIC ANALYSIS ······ Quality: GOOD

✓ SECURITY AUDIT ····· Risk Level: HIGH

✓ PERFORMANCE REVIEW · Efficiency: FAIR

✓ DOCUMENTATION REVIEW Coverage: FAIR

── AGGREGATING RESULTS... ──────────

Score 78/100 · Verdict: NEEDS MINOR WORK

1. remove hardcoded secret in config.py

saved → output/reports/review_src_<timestamp>.md

06 AI Agents · Developer Tools

Autonomous Multi-Agent Code Reviewer

A CLI that reviews code from four angles at once. Four specialist agents run concurrently, then a fifth aggregator merges everything into one prioritized report with a quality score.

Four reviewer lenses: StaticAnalyzer (quality, anti-patterns), SecurityAuditor (OWASP Top 10, injection, secrets, CVE lookup), PerformanceReviewer (complexity, N+1, async misuse) and DocChecker (docstrings, type hints, dead code).
All four run with asyncio.gather, so a full four-angle review takes about as long as a single agent call.
A ReportAggregator agent deduplicates overlapping findings, ranks by severity and produces a 0–100 score, verdict, executive summary and top 3 actions, enforced with Pydantic schemas.
Prints each reviewer’s report, then the aggregated report, and saves a Markdown report; supports Python, JS, TS, JSX/TSX, HTML, CSS, JSON and YAML.

4 + 1Parallel reviewers + aggregator

0–100Score, verdict & top 3 actions

Python
OpenAI Agents SDK
asyncio
Pydantic
Brave Search API

View code

More projects

AI · Computer Vision

Intelligent Waste Classification System

A waste recognition web app: upload a photo and a ResNet50 model built in PyTorch classifies the material into one of 7 categories (plastic, glass, paper, metal, cardboard, cloth and other) with 92% accuracy, then shows recycling guidance for it.

PyTorch
ResNet50
Django
SQLite

Code

Attendance system recognizing a face in a live camera frame and labeling it with the student's name

AI · Computer Vision

Facial Recognition Attendance (Prototype)

A prototype attendance system that recognizes students' faces in real time and marks attendance automatically, with separate student and teacher logins, attendance reports, and checks against proxy attendance.

Python
Django
face_recognition
Computer vision

Code Demo

04 / Research & Awards

Research & recognition.

Adversarial machine learning research at Georgia State, and building under pressure at hackathons.

Papers & thesis

ManuscriptIEEE format · Georgia State University

Defending End-to-End Face Authentication Pipelines via Internal-Activation Feature Fingerprinting

C. Onwuchuluba, C. Killari Giribabu
- 0.997 AUROC
- 1,411-d fingerprint
Course paperCSC 8230 · Georgia State University · Spring 2026

Adaptive Multi-Stage Defense Against Adversarial Examples: Progressive Intensity Matching Input Difficulty

C. Onwuchuluba, C. Killari Giribabu
- 85–93% detection
- 6.4 avg forward passes
In progressGeorgia State University

M.S. Thesis

Thesis-track Master of Science in Computer Science

HackATL 2025 Finalist

PulseGuard

A predictive safety wearable built in 48 hours. It runs an ML risk model on sensor data to detect hazardous chemical exposure in real time, with an automated alerting dashboard that flags at-risk users before exposure becomes dangerous. Built with team LunaTek.