Projects
27 projects across computer vision, applied ML, full-stack products, developer tooling and research.
Page 1 of 3
AVZDAX Behavioural Detection Platform (Visual Intelligence)
CCTV behavioural-analytics platform flagging loitering, tailgating, clustering and erratic movement across retail and estate camera feeds, with a companion object-ownership tracker for abandoned and stolen items.
Status: Internship work (July 2025 – Present)
Role: Work Experience (Internship Project)
Key Features:
- Built the detection stack on YOLOv8 nano and beta YOLO11 nano for person and pose detection, with the behaviour rules hand-written in NumPy — the clustering detector builds an…
- Wrote a monocular depth module using a pinhole model (known object height divided into pixel height, times focal length), plus a scale normaliser and distance-invariant features…
- Tuned a BoT-SORT tracking configuration with re-identification from the detector's own features, holding one ID per person through occlusion without shipping a second model file.
Obex Edge Inference Service (AVZDAX)
Edge inference process that runs computer-vision detection on a Raspberry Pi attached to one or more RTSP cameras and raises alerts — annotated clips plus metadata — to the Obex backend for storage, notification fan-out and review.
Status: In Progress (2026)
Role: Work Experience (Internship Project)
Key Features:
- Decoupled capture from inference through Redis: a stream manager parks frames with a short expiry and the analyzers consume at their own pace, so a slow model cannot stall the…
- Built six detectors — weapon, aggression, shoplifting, intrusion, loitering, and queue intelligence with checkout bypass — switched on per stream by database flags, with each…
- Wired three ONNX models trained externally for the edge: a knife/handgun object detector and two image classifiers for aggression and shoplifting.
RTLT (Real-Time Language Translation)
Low-latency, real-time speech-to-speech voice translation system providing bidirectional English ↔ Yoruba translation.
Status: In Progress (April 2026 – Present)
Role: Research Collaboration (NITHUB)
Key Features:
- Collaborating with NITHUB AI researchers to design low-latency WebSockets audio streaming channels.
- Integrating Meta SeamlessM4T v2 and Silero VAD for voice activity detection and translation quality evaluation.
Oloja
AI-powered gig matching and financial identity platform for informal workers in Nigeria. Replaced manual hiring with automated AI verification.
Status: Completed (Demo/Top 50 candidate)
Role: Hackathon (Squadco Hackathon 3.0, Challenge 02)
Key Features:
- Reached the Top 50 round (ranked 3rd stage, fell short of the final Top 10 pitch).
- Fully integrated a NestJS backend with Claude Haiku 4.5 for semantic matching of worker skills (e.g. plumbing, tiling) to opportunities.
- Implemented an algorithmic Trust Score weighting volume (transaction count), consistency, recency (exponential decay), and reliability to unlock micro-savings and credit.
Prune Docs (Duplicate & Similar Document Finder)
Desktop app that finds duplicate and near-duplicate documents across a folder using Qdrant vector search over Superlinked schemas, with an Electron front end. This is the built version of the earlier Qdrant hackathon concept.
Status: Completed (2025)
Role: Hackathon Project (Qdrant Hackathon)
Key Features:
- Superlinked schema and index declaring text, file-path, file-name and job-id similarity spaces plus categorical type and recency spaces over Qdrant.
- FastAPI service loads the Qdrant executor once at startup and exposes /store-documents and /prune-documents.
- Batched embedding ingestion under a per-job UUID, with a Qdrant payload index on job_id and grouping at a 0.9 similarity threshold.
VIGIL ("Your Estate Never Sleeps")
Visual intelligence for gated estate security and access management.
Status: Completed (Demo - Collaboration with Akorede, Charles, and Peace)
Role: Hackathon (Claude Code Hackathon)
Key Features:
- Real-time resident face + plate verification to open gates.
- Visitor verification flow requesting host approval via push notification.
- Scene threat detection (perimeter climbing, fire, medical emergencies).
Chess Analysis & Bot Tournament
Local chess training and bot tournament pipeline evaluating models against Stockfish.
Status: Concept/Experimental
Role: AI / Machine Learning Study (Personal)
WhotAI
Python-based simulation of the classic Nigerian card game Whot! featuring custom rules and simulated AI card players.
Status: Completed
Role: Game Development / AI Study (Personal)
Key Features:
- Implemented core card game rules: PICK_THREE, SUSPENSION, DOUBLE_DECKING, and COMPOUND_DEFENSE.
- Built StatelessAIWhotPlayer powered by a PyTorch MLP (Multi-Layer Perceptron) neural network to evaluate game states and rank valid card plays.
Transactional Intelligence for SME Continuity (Heirs Insurance Hackathon 2026)
AI-driven middleware platform for real-time stock value estimation and continuous underwriting for informal SMEs, facilitating instant claims processing.
Status: Completed (Submitted February 2026)
Role: Hackathon (Team GIG -- University of Lagos)
Key Features:
- Integrated financial data inflows/outflows via transaction APIs (Moniepoint, OPay, Flutterwave) to compute real-time digital proof of stock.
- Built a "Listen and Learn" continuous underwriting algorithm that dynamically adjusts the Sum Insured to protect against hyperinflation.
- Formulated a Pay-as-you-Stock (PAYS) daily micro-premium structure to make coverage affordable for low-margin traders.
Food Pricing Causal Inference & Forecasting
A causal inference and time-series forecasting pipeline designed to predict and explain food commodity prices in Nigeria.
Status: In Progress (April 2026 – Present)
Role: Work Experience (NITHUB Internship Project)
Key Features:
- Built robust data ingestion pipelines to align and structure highly varied inputs including meteorological data, NGN/USD exchange rates, and historic commodity prices.
- Utilized statistical forecasting models to identify major economic causal factors driving local food inflation.
Polynomial Root Solver (SSG 226)
An interactive Python command-line utility implementing numerical analysis algorithms to solve higher-degree polynomials.
Status: Completed (July 2025)
Role: Academic Project (University of Lagos)
Key Features:
- Implements Newton-Raphson and Regula-Falsi (False Position) methods for finding real roots of arbitrary degree polynomials.
- Features robust error handling, decimal precision parsing, and runtime input validation loops.
- Dynamically renders polynomials and intermediate iterations in mathematical notation.
Biogen (AI Cancer Risk Assessment)
Multimodal early risk assessment AI fusing breast/cervical imaging data and clinical biomarkers to predict 1- to 3-year cancer risk.
Status: Discontinued / Abandoned (February 2026)
Role: Hackathon / Collaboration
Key Features:
- Implemented an explainability layer providing clinicians with visual heatmaps and feature importance metrics.
- Designed a clinician-facing frontend to review patient risk profiles and treatment signals.