I build AI systems that do real work under real governance — agents that prepare, humans who decide, and evidence for everything. The most valuable thing an AI system can do is refuse: to guess, to improvise, to pretend it knows.
- 🏛️ Flagship: TaxOS — an open-source enterprise agentic tax operating system. A multi-agent runtime carries UK VAT and Corporation Tax from ERP extract to a review-ready, evidence-attached state — and architecturally cannot approve its own work. Deterministic rule engine with HMRC-cited lineage, hash-chained audit, one-click evidence packs, RAG that cites or refuses, explainable ML that declines to train without enough labels. 169 tests · 18 ADRs · runs locally, no API keys
- 💼 Currently: AI Engineer at Blockchain Advisors — owning the AI capabilities of a live AI-enabled CRM platform (RAG chatbot with human hand-off; vision pipeline behind an admin review queue), ex-PwC Tax Technology
- 🧠 Focus areas: agentic systems & LLM applications (RAG, fine-tuning, safety pipelines), explainable ML, medical imaging & NLP moderation
- 📊 I care about honest evaluation — calibration, thresholds, and knowing when a metric is lying to you
- 🔬 Also shipped: SkinSense (clinical triage + RAG assistant) and Atlas AI, a private-markets analytics platform (private repo — happy to discuss)
Languages & Core
ML / AI
Backend & Frontend
Deployment & Tools
| Project | What it is | Stack |
|---|---|---|
| 🏛️ TaxOS — Enterprise Agentic Tax Operating System | Governed multi-agent platform carrying UK VAT + Corporation Tax from ERP extract to a review-ready, evidence-attached state: deterministic rule engine with HMRC-cited per-figure lineage (tax types as content packs — CT added with zero engine changes), an approval gate agents architecturally cannot reach, hash-chained audit with one-click evidence packs, RAG that cites or refuses, and explainable ML that declines to train without labels. 169 tests · 18 ADRs · no API keys | FastAPI · PostgreSQL · Next.js · scikit-learn |
| 🔬 SkinSense — AI Skin-Lesion Triage | Vision triage model (EfficientNet-B0, INT8, temperature-scaled, RL-adaptive threshold) + DermBot, a Gemini-powered clinical assistant grounded in 8,700+ clinical documents via RAG, with a multi-layer safety pipeline | PyTorch · FastAPI · React · Gemini · RAG |
| 🛡️ Multilingual Hate-Speech Moderation | Cross-lingual moderation system — XLM-RoBERTa + LoRA fine-tuning with an RL policy that outputs actionable decisions (PASS / WARN / FLAG / REMOVE). Live demo → | Transformers · LoRA · RL · FAISS · React |
| 💳 Credit-Card Fraud Detection | An honest case study in imbalanced classification: why 95.5% recall can hide a coin-flip model (ROC-AUC ≈ 0.51), and how to diagnose it | scikit-learn · imbalanced-learn · Jupyter |
| 📈 Time-Series Forecasting | Decomposition, stationarity testing, and ARIMA-family forecasting notebooks | statsmodels · Pandas |
- 📧 Email: olisaanthony25@gmail.com
- 💼 LinkedIn: linkedin.com/in/anthony-onochie-olisa-17031021a
- 🤗 Hugging Face: huggingface.co/Stoic1344223
Open to AI/ML engineering roles and collaborations.