Patrick Adegbesan

Professional Summary

Full-stack and AI engineer who takes systems from architecture through to deployment and operation. Builds multi-agent LLM pipelines, reinforcement-learning systems, and production web platforms in Python, TypeScript/React and Node.js — together with the operational layer that keeps them running: Docker and systemd deployment, scheduled retraining, alerting, monitoring dashboards, and written runbooks. Delivers client work end to end, from requirements and UX through to a live, maintained product, and works AI-native day to day with Claude Code, Claude AI and Claude Cowork as primary tools.

Technical Skills

• Languages: Python, TypeScript, JavaScript, SQL, PHP, Java, HTML5, CSS3
• AI & LLM: Claude (Code, Cowork, API), OpenAI, Gemini, Kimi/Moonshot, LangChain, multi-agent orchestration, RAG, prompt & context engineering, LLM evaluation (ROUGE-L, BERTScore, NDCG)
• Machine Learning: PyTorch, scikit-learn, deep reinforcement learning (DQN, Double/Duelling, Prioritised Experience Replay), spaCy, MediaPipe, DeepFace, feature engineering, backtesting
• Vector & Search: ChromaDB, FAISS, embeddings, semantic retrieval
• Frontend: React, Next.js 14, React Native/Expo, Vue, Tailwind CSS, Radix UI, Material-UI, Vite, responsive & accessible UI (WCAG, ARIA)
• Backend: Node.js/Express, FastAPI, Flask, Django + DRF, REST API design, JWT/OAuth, NextAuth, webhooks, WebSockets
• Databases: PostgreSQL, MySQL, MongoDB, SQLite, Prisma ORM, schema design, row-level security
• Infrastructure & DevOps: Docker & Docker Compose, systemd services and timers, AWS S3, Stripe/Stripe Connect, Vercel, Render, Netlify, GitHub Pages, Git/GitHub, CI/CD
• Operations: Runbook authoring, monitoring & alerting (Telegram bots, watchdogs), scheduled retraining with promotion gates, audit logging, incident playbooks, risk controls & circuit breakers
• Data & Analytics: Data cleaning, EDA, visualisation (Chart.js, matplotlib), Excel/Google Sheets, reporting pipelines

Experience

• Software Engineer — 3O1Atech (2026 – Present):
  • Build and maintain CampusDash, a four-surface campus platform spanning a customer web app, a cross-platform mobile app, a backend API service and an internal admin dashboard — each its own repository with a shared contract between them
  • Delivered GhostGuard at Squad Hackathon 3.0: a continuous AI payroll-integrity layer that re-verifies worker identity every cycle before funds are released, gating on liveness detection (MediaPipe FaceMesh), face matching (DeepFace Facenet512) against a BVN reference, multi-signal deepfake analysis, and behavioural anomaly scoring over transaction history
  • Integrated the Squad payments API end to end — BVN-validated virtual accounts as identity anchors, gated transfers, account lookup, and HMAC-validated webhooks into a tamper-evident audit log built as a prosecution-ready evidence package
  • Built a 14-agent LLM system for the DSN × BCT Hackathon 3.0: a six-agent chain for persona-calibrated review generation and an eight-agent cross-domain recommender, behind a pluggable LLM backend (Kimi, Gemini, OpenAI, local HF) with a ChromaDB vector store
  • Wrote the evaluation harness for it — RMSE, ROUGE-L, BERTScore, NDCG and Hit Rate — so model changes were measured against a baseline rather than judged by eye
  • Shipped the company website at 301atech.com
• Independent AI & Systems Engineering (2025 – Present):
  • Project Cognition — ensemble reinforcement-learning trading system carried from architecture document to deployable service across seven milestones: five DQN agents voting through independent feature lenses, with a per-regime meta-learner that reweights them after every closed trade
  • Designed a standalone risk engine that every order must clear before execution and that agents cannot override — 1% equity cap per trade with fractional-Kelly sizing, a −5% daily circuit breaker, a drawdown governor, overtrading guards, and config validators that refuse any value looser than the hard limits at load time
  • Built the operational layer for unattended 24/7 running: Docker Compose with health checks, systemd units and timers, nightly retraining behind a promotion gate that refuses to activate a worse model, Telegram alerting with throttling, a feed watchdog, a self-contained HTML dashboard, and a written runbook covering deployment, monitoring, failure playbooks and rollback
  • Built a nine-system software archive — each system self-contained, offline-first with vendored assets, individually documented, and shipping a terminal client with full parity to its web UI, driving the same code and database
  • That archive includes PhishGuard AI (32-feature ML phishing detection), SentinelIQ (credit-card fraud scoring with a live analytics dashboard), an automated resume screener doing PDF/DOCX parsing, skill matching, geospatial proximity ranking and Excel export, and TaskMind (plain-English task capture with priority scoring and next-action recommendations)
  • Academia Hub — two-sided academic marketplace on Next.js 14 with Stripe Connect seller onboarding and split payments at a 15% platform commission, webhook-tracked payment state, AWS S3 storage behind signed URLs, buyer–seller messaging, and role-based access across buyer, seller and admin
  • CardExpress — ATM card request and delivery system for Nigerian banks covering the full request-to-delivery lifecycle, with public reference-code tracking that needs no login and an accessibility-first build (keyboard navigation, ARIA live regions, persistent high-contrast mode) in service of its financial-inclusion goal
  • Hospital EHR System — Django + DRF and React records platform enforcing strict row-level security, where only doctors on a patient's authorised list can reach their data and access is shared explicitly by doctor ID or email
  • Adaptive RL research — enhanced DQN with Double Q-learning, duelling architecture and Prioritised Experience Replay, trained over ~1,122 episodes and evaluated zero-shot across seven altered-physics environments, written up as an academic paper
  • AI PDF Reader Bot — Streamlit chatbot over multiple PDFs using LangChain, OpenAI and FAISS retrieval, with text chunking strategies and conversation memory for contextual Q&A
• Freelance Developer — Client Delivery (2025 – Present):
  • Fortune et Feveur (tiling company) — built the F&F Inventory Management System, a cross-platform Electron desktop app with a React frontend and Node.js backend handling inventory tracking, sales processing, reporting and receipt printing, over ~92 commits; also delivered and deployed the public site at fortuneetfeveur.com
  • Palmwine Merchants — event booking and ticketing platform for event discovery, online ticket purchase and brand engagement, supporting recurring event operations
  • SLO — translated a Figma design system into a working React + TypeScript brand site
  • Ran each engagement end to end: requirements gathering, direct client communication, delivery, deployment and ongoing maintenance

How I Work with AI

• Claude Code as a daily driver — primary tool for building, refactoring and shipping. This portfolio, including its content and structure, is maintained through Claude Code sessions against the live repository.
• Claude AI for research and decision support — deep research on tools and approaches, synthesis into short actionable summaries, and drafting technical documentation and client-facing communication.
• Claude Cowork for documentation and execution — organising project knowledge, keeping written context current, and moving multi-step work forward.
• Building with LLMs, not just using them — multi-agent orchestration, RAG over vector stores, provider-agnostic backends with graceful fallback, and measured evaluation rather than vibes.
• Automation bias — when a task repeats, it becomes a script, a CLI or a scheduled job. Every system above ships with its own tooling for exactly this reason.

Certifications

Cisco Networking Academy — Data Analytics Essentials (Issued: May 6, 2025)

The Code Center — AI & Software Development (Issued: Nov 25, 2025)