I build AI into products end-to-end — from LLM pipelines and agent systems to the backends, mobile apps, and infrastructure that run them in production. Built a GPT from scratch in PyTorch. Fine-tuned BERT/T5/Mistral across 6 domains. Designed autonomous multi-agent systems before LangChain existed.
📹 Watch Portfolio Loom (90s) · 🐙 github.com/johnmoses · 🌐 shopstack360.com
10,000+ users in week one · FastAPI + Flutter + Next.js · AWS Lightsail · $20/month
A full-stack attendance reporting platform for hierarchical church organizations (Fellowship → Zone → District → Region → State). Built offline-first — works without internet, syncs when connected. Security: guest, scoped event worker, and 3-tier admin roles enforced via FastAPI RBAC at the dependency injection layer.
AI highlights:
- Unified Chat — LangGraph supervisor + 6 specialist agents in one compiled StateGraph. Single interface for report submission and data querying in natural language. Supervisor classifies intent, report agent does multi-turn field collection, 5 query agents (insights, list, forecast, comparison, missing) each render to text / WhatsApp / PDF / Excel. Redis-backed session with in-process fallback.
- Voice capture pipeline — Whisper (real-time transcription) → GPT-4o-mini (structured field extraction via function calling) → regex fallback running in parallel. Offline-first with server-side recovery on reconnect. Handles Nigerian multilingual input (Hausa, Yoruba, Igbo, Efik, Idoma)
- Attendance anomaly detection — flags consecutive member absences and center-level attendance drops across all centers automatically
- Outreach intelligence — engagement scoring, churn prediction, and message variant generation by tone
- Statistical trend analysis — Redis-cached insights with cold-start handling for new centers
DevOps highlights:
- ADR-driven deployment decisions documented before any infrastructure was touched
- Docker Compose with per-service memory limits, PostgreSQL tuned for a $20 Lightsail instance
- GitHub Actions CI/CD, Nginx reverse proxy, fail2ban hardening, kernel-level SYN flood protection
- Sliding window rate limiter with real IP extraction behind proxy
Code excerpts:
- Voice capture pipeline — hybrid inference, multilingual NLP, offline resilience
- Anomaly detection agent — autonomous absence + drop detection
- Offline-first sync — Flutter + Next.js dual-layer queue
- Production infrastructure — Docker, Lightsail, hardening, rate limiting
- Outreach AI — engagement scoring, churn prediction, priority queue
Pre-AI-era foundation (built without AI assistance):
- Custom transformer fine-tuning — BERT intent classifier (44 classes, 87.5% accuracy) + T5 text-to-SQL (ROUGE-L 0.961 PEFT/LoRA variant, 118K samples) · 3 models published to HuggingFace · all runs tracked in Weights & Biases
- Conversational agent — multi-turn dialog state machine with entity extraction and autonomous DB writes via WebSocket (agentic pattern before LangChain existed)
Co-Founder & Lead Engineer (team of 5) · FastAPI + Flutter + Next.js · PostgreSQL + Redis · Docker · Railway
A full-stack B2B/B2C commerce platform connecting manufacturers, distributors, retailers, and shoppers across Nigeria and West Africa. Incubated in Abuja with 20+ live retailers, scaling to the Lagos industrial corridor.
AI highlights:
- 5-agent autonomous fleet — Smart Reorder, Supplier Intelligence, Price Optimization, Inventory Advisor, Product Recommendations — orchestrated via LangGraph with an eval harness (pass rates persisted to Redis) and an intervention tracker (human override rate + trend per agent per 7-day window)
- Multi-provider AI strategy — pluggable provider architecture (rule-based, LangChain hybrid, AWS Bedrock) activated by feature flags; zero downstream changes when provider switches
- Production MCP server — wraps the agent fleet as typed tools (stdio + HTTP transport); open-source mcp-foundations repo (5 progressive apps) demonstrating MCP protocol from basic FastMCP through full-stack agent bridges
- Arbitrage engine — 3-layer price scanner (internal + market scouts + external), landed cost calculator (product + shipping + customs + last-mile), opportunity scorer (0–100 composite)
- Borderless logistics AI — route optimization, rider matching, demand forecasting (hourly heatmaps), dynamic surge pricing
- Social intelligence signals — tier-specific feed items generated from live transaction data; agent health signals surface in business-tier feeds
Cloud & infrastructure:
- Lightsail is the stable core — 7 services (backend, AI, web, DB, Redis, MCP, autoheal) run on Docker Compose with self-managed equivalents for every AWS capability; the platform never depends on a managed service being available
- ECS is a detachable burst layer — containers are ECS-compatible by design (health endpoints, stateless, stdout logging), but ECS only attaches for heavy workloads (model retraining, batch ML inference, large arbitrage scans); if ECS has a cold start, task failure, or cost spike, Lightsail handles the same workload in degraded-but-running mode
- Lambda sits between them as a pure scheduler — fires the right target (ECS task or Lightsail endpoint) based on availability; owns no business logic
- Bedrock, RDS, and S3 follow the same pattern: each has a Lightsail-side fallback (local inference, self-managed Postgres, local artifacts); switching is a feature flag, not a rewrite
- 3-layer observability: infrastructure metrics + model quality (eval harness) + data drift detection (TFDV) as retraining signal
- 24h auto-retraining loop — reloads recommendation engine in-place on success, keeps previous version on failure
Full-stack highlights:
- 3-provider payment routing (Paga / Paystack / Flutterwave) — intelligent routing saves 45% on processing costs
- Escrow with milestone releases for B2B, wallet system for instant settlements
- 7-level ambassador engine with anti-fraud detection and dormancy downgrade
- Community commerce: group buy lifecycle with per-participant escrow and trust scoring
Private repo — architecture overview.
| Repo | What it covers |
|---|---|
| llm-foundations | LLM internals, fine-tuning, prompt engineering, inference optimization |
| rag-foundations | RAG pipelines, vector stores, chunking strategies, retrieval evaluation |
| ai-agents-foundations | Agent loops, tool use, multi-agent coordination, autonomous control |
| mcp-foundations | Model Context Protocol — 5 progressive apps, stdio + HTTP transport, full-stack agent bridges |
| Repo | What it covers |
|---|---|
| ultra-learning | AI-powered adaptive learning system — Flask API (LLM + RAG + MCP) · Next.js web · React Native mobile |
| sure-health | Multi-party AI healthcare platform — Flask API (agents + LLM + RAG) · Next.js web · React Native mobile · 90%+ FHIR compliance |
| easy-finance | AI-powered financial platform — FastAPI (wealth + community + security services) · Next.js web · React Native mobile |
| zero-to-hero-python-ai | End-to-end ML/AI roadmap — CNNs, RNNs, Transformers, LLMs, Agentic AI |
| Repo | Credential | Score |
|---|---|---|
| coursera-genai-agents-specialization | Building GenAI Applications and Agents (6 courses) | All 100% |
| coursera-nlp-specialization | Coursera NLP Specialization | C1–C3 100%, C4 97.5% |
| coursera-mlops-specialization | Coursera MLOps for Production | C1–C3 100%, C4 98% |
| coursera-ai-4-medicine-specialization | Coursera AI for Medicine (3 courses) | All 100% |
| Repo | What it covers |
|---|---|
| jenkins-docker-compose-flask | Jenkins CI/CD with Docker Compose |
| jenkins-docker-flask | Jenkins + Docker pipeline |
AI/ML: LLMs · SLMs · RAG · Agentic AI · MCP · LangGraph · AutoGen · CrewAI · Transformer Fine-Tuning (BERT, T5, GPT-2, Mistral) · PEFT/LoRA · QLoRA · Collaborative Filtering · Time-Series Forecasting · NLP · Computer Vision · Weights & Biases
Backend: FastAPI · Flask · PostgreSQL · Redis · pgvector · Qdrant · SQLite · REST · WebSockets · SSE
Frontend: Next.js · TypeScript · Tailwind CSS · React
Mobile: Flutter · Dart · React Native · Offline-first (SQLite + sync queue)
DevOps & MLOps: Docker · GitHub Actions · AWS Lightsail · ECS · CloudWatch · Railway · Nginx · fail2ban · TFDV · TFX
📧 johnmosesng@gmail.com · 🐙 github.com/johnmoses · 📹 Loom Portfolio · 🌐 shopstack360.com
30 courses across 5 institutions — 22 at 100%, 7 above 95%.
| Area | Credential | Score |
|---|---|---|
| GenAI & Agents | Building GenAI Applications and Agents Specialization (6 courses) | All 100% |
| AI for Medicine | AI for Medicine Specialization (3 courses) — DenseNet/U-Net, Cox PH, GradCAM | All 100% |
| MLOps | MLOps for Production Specialization (4 courses) | C1–C3 100%, C4 98% |
| Deep Learning | Deep Learning Specialization (5 courses) | C3 & C5 100%, others 97%+ |
| NLP | Natural Language Processing Specialization (4 courses) | C1–C3 100%, C4 97.5% |
| Mathematics | Mathematics for ML: Linear Algebra · Multivariate Calculus · PCA · Data Science Math Skills | All 98%+ |
| Finance/ML | Investment Management with Python & ML Specialization (4 courses) | 3 courses 100%, 1 at 97%+ |
| Generative AI | Generative AI with LLMs | — |