Building

Payment escrow trust layer for Nigerian freelancers and clients. AI milestone structuring assistant in Phase 1; agentic dispute resolution and risk scoring in Phase 2. Tagline: "Your payment is safe. Your work is paid."

Production-grade ML model serving via FastAPI with containerized deployment and a statistical drift detection layer. Paired with an automated CI/CD pipeline as Part 1 of the ML Production Stack.

Automated ML lifecycle — pytest API tests → model evaluation → gated deployment via GitHub Actions. Model only ships if it clears a 85% accuracy threshold. Part 2 of the ML Production Stack alongside the drift monitoring API.

Fine-tuned Mistral-7B on a curated Pan-African fintech dataset using LoRA + 4-bit quantization on a free T4 GPU. Only 1.11% of parameters trained. Benchmarked with ROUGE evaluation — ROUGE-1: 0.47 on domain-specific content with 18 training samples. Adapter live on HuggingFace Hub.

Three-agent CrewAI system — Planner, Researcher, Writer — that autonomously researches African fintech topics and produces structured market intelligence briefings. Deployed on Cross-border stablecoin payments in West Africa as the first research run.

A content strategy agent that remembers past Kemdai campaigns using Supabase pgvector episodic memory. Pulls recent and semantically similar past content, injects it as reasoning context, and suggests fresh angles that avoid repetition. Memory compounds over time — the more you use it, the smarter it gets.

A production ML pipeline that monitors its own performance and autonomously fixes itself. Five LangGraph agents — Monitor, Diagnostic, Retrain, Evaluate, Promote — coordinate to detect drift, diagnose root cause via Claude, retrain, benchmark, and swap the live model without restarting the server. Every run logged to MLflow.

Real-time AI intelligence platform monitoring fintech news across Nigeria, Kenya, Tanzania, Ethiopia, and the DRC. Articles are ingested from NewsAPI, classified by Claude via LangChain, and surfaced on a React dashboard with market, category, sentiment, and confidence scoring. MLflow tracks pipeline runs; graceful degradation handles cloud-tier constraints.

Two GAN applications under one experimental ML project: AfroCover generates African-inspired album artwork via StyleGAN2-ADA, and Lagos2Duplex transforms old Lagos houses into modern duplex mockups via CycleGAN/CUT. Standardized FID/KID/LPIPS evaluation, CPU-friendly training, Docker/K8s deployment, and a live Gradio demo on Hugging Face Spaces.

Production-ready RAG that prioritizes measurable correctness. Hybrid retrieval (vector + keyword fused with RRF), LLM grounding with citations, and a built-in evaluation layer scoring faithfulness and usefulness. A failure taxonomy classifies retrieval misses, hallucinations, and over-refusal — and a re-query agent attempts alternative phrasings when confidence is low.

An economically-active AI agent that hunts for mispricings on Polymarket. Scans 50 high-volume markets every 30 minutes and scores each opportunity 1–10 via Gemini 2.0. Anything scoring 7 or higher triggers a Telegram alert and gets posted as analysis to Moltbook. Includes a paper-trading ledger, peer-review messaging across agents, and a heartbeat module for persistence.

Production news recommendation platform that ingests RSS feeds every four hours, embeds articles with a HuggingFace inference endpoint, and ranks them per-user via vector similarity. Maximal Marginal Relevance (MMR) keeps the feed diverse and avoids filter bubbles; trending injection keeps breaking news visible. JWT auth + live deployment on Render.

End-to-end weather ETL plus an XGBoost classifier that predicts rainfall 6–12 hours out. Isotonic calibration produces well-behaved probabilities; SHAP explains every prediction; dual operating modes (recall-focused vs precision-focused) let users pick their tradeoff. Fully reproducible via Makefile, served through FastAPI and a Gradio demo.

Fraud-detection ML stack combining Logistic Regression and LightGBM over 25+ behavioral, temporal, and deviation-based features. Temporal-safe train/test splits prevent leakage; PR-AUC optimization handles severe class imbalance; threshold tuning prioritizes precision to keep false-positive cost low. Served via FastAPI with batch + single-record endpoints and Docker deployment.

Annual rental price estimator for Lagos mainland properties. Multi-source scraping pipeline feeds a scikit-learn regression model with Box-Cox normalization; serialized artifacts auto-download from the Hugging Face Hub when missing locally. Gradio UI returns a price plus a list of comparable listings, with smoke tests guarding CI.