Tiamiyu – LLM, Machine learning, Python

A fintech platform focused on automating financial operations through AI-powered workflows for accounting and finance teams. The platform includes a multi-agent bank reconciliation engine that automates transaction matching and reconciliation processes, achieving 95% matching accuracy and reducing manual effort by 90%. It also provides an automated document processing pipeline for invoices and financial statements, streamlining data extraction and processing while reducing operational overhead by 93%.

• Architected a LangGraph multi-agent bank reconciliation system that eliminated 90% of manual reconciliation work and hit 95% matching accuracy across thousands of monthly transactions. Replaced a process that previously consumed days of analyst time per cycle.
• Designed a SQL agent with a human-in-the-loop feedback loop. Analysts correct bad queries and those corrections feed back into the system, improving accuracy on ambiguous financial data over time without retraining the model.
• Built a multi-provider LLM evaluation framework that benchmarks several proprietary and open source models across 28 analyst agent test cases using deterministic scorers and LLM-as-judge, with automated benchmark rankings, head-to-head analysis, and LLM reports. Enabled data-driven model selection for production.
• Built an offline evaluation framework that benchmarks document extraction accuracy across 3 OCR engines and multiple LLM models using normalized exact match, LLM-as-judge, hallucination detection, and 4 other automated metrics. Enabled data-driven OCR and model selection for production.
• Led a data governance assessment of the AI platform, tracing sensitive data flows across 13 third-party services and identifying 75 compliance gaps related to encryption, PII exposure, and data protection regulations. Proposed remediation strategies across three priority levels.
• Owned the full AI architecture from problem definition to production deployment, working without PM oversight and translating business requirements directly into system design.

A payroll fraud detection system that evaluates every employee across five detection layers before payroll files are released, enabling proactive fraud prevention rather than post-payment auditing. The platform integrates live identity verification (NIN/BVN) with ERP attendance data and provides plain-English explanations for each fraud verdict, supported by a complete audit trail and frontend dashboard. In a live test across a 40-person payroll worth ₦11.48M, the system identified and blocked 15.2% of the total payroll value at risk.

• Built a 4-agent system (Verifier, Auditor, Guardian, Red Team) that detects ghost workers before payment runs. In a live test of 40 employees, it protected an estimated 7.92M naira in annual exposure, equivalent to twelve months of undetected fraud stopped on the first scan.
• Designed a deterministic 5-layer detection engine (identity, shared attributes, existence, process, cross-check) that scores every employee and produces explainable verdicts. Blocked 3 hard-fraud cases and flagged 15.2% of total payroll value in the demo run.
• Built a SHA-256 hash-chained audit trail where database triggers reject UPDATE and DELETE at the storage layer. Every verdict is reproducible and traceable to a named rule and a piece of evidence, regardless of who has application-level access.
• Integrated live external identity verification (Dojah NIN/BVN/NUBAN), Odoo ERP for HR and attendance data, and a fallback chain (Gemini -> Llama 3.3 70B -> rules-only) so detection keeps running even when third-party APIs go down.
• Built a Red Team agent that attacks the system's own defenses with 5 scripted adversarial scenarios on every run. Currently catching 5 of 5.
• Shipped a working demo in under 72 hours by using Claude Code and Cursor for scaffolding, while hand-engineering the core detection engine, hash-chain audit trail, and agent orchestration. Cut sprint time significantly without touching the architectural fundamentals.

A legal assistance platform that analyzes tenancy agreements in under 10 seconds, identifying potential rights violations and predatory clauses to help tenants better understand their contracts before seeking professional legal advice. The system supports four languages with full voice input and output, making legal document analysis more accessible across different users and language preferences. It combines a retrieval-augmented generation (RAG) pipeline with multilingual voice capabilities and a full-stack application, providing fast, accessible contract analysis at a fraction of the cost of a traditional legal consultation.

• Built a legal RAG system combining a regionally-accented LLM (N-ATLaS) with hybrid vector search (dense embeddings + sparse BM25) in Qdrant. Hits under 10 seconds for full tenancy agreement analysis at 95%+ accuracy in identifying predatory clauses.
• Shipped voice-first interaction by integrating YarnGPT TTS and N-ATLaS ASR across 4 languages. Users who cannot read English can speak their questions and hear answers back in their own language.
• Architected a multilingual RAG pipeline that holds conversation context across document uploads, streaming responses, and language switches in a single session without dropping retrieval quality.
• Delivered a full-stack application (FastAPI + React + Qdrant + Modal) deployed to Google Cloud Run and Vercel, with a freemium B2C model and a B2B API licensing path built into the architecture from the start.

Healthcare claims processing platform for hospital billing teams. It accepts multi-document patient encounters (lab reports, pharmacy logs, billing receipts), extracts structured claim fields using a language model, and moves each claim through a 7-stage review workflow until HMO submission. I built the full system, from the AI extraction pipeline and async task queue to the React frontend and deployment setup, as a capstone project.

• Built end-to-end document extraction using Google Gemini that pulls structured NHIA claim fields (patient ID, diagnosis codes, tariff codes, costs) from unstructured multi-document patient encounters. Replaced a fully manual data entry process.
• Engineered an async pipeline with Celery and Redis that decouples document upload from extraction. Multiple billing officers can process claims concurrently without blocking each other or losing state mid-workflow.
• Implemented a stateful 7-stage claim lifecycle (DRAFT -> PROCESSING -> EXTRACTED -> UNDER_REVIEW -> READY -> SUBMITTED -> PAID/REJECTED) with role-based review gates. Standardized a previously ad-hoc HMO submission process and reduced claim rejection risk.
• Delivered a full-stack system (React + FastAPI + PostgreSQL) containerized with Docker Compose, deployable to Railway with a single environment file. Built specifically for hospital IT environments with limited infrastructure.

A data platform for large-scale entity resolution and analytics. I built the ML systems for matching records across inconsistent datasets and giving non-technical users direct access to structured data through natural language queries, replacing slow manual processes on both fronts.

• Built a record linkage system using sentence embeddings that resolved entities across noisy, inconsistent datasets with high precision. Replaced fragile rule-based string matching that was breaking on real-world data quality issues.
• Implemented a text-to-SQL pipeline using Vanna-AI that lets non-technical users query large structured datasets in plain English. Cut analyst query turnaround from hours to seconds and removed the SQL bottleneck from the data team.