Results · five real systems

Five agentic systems, built and running.

Most new AI studios can't show a single working agentic system. I'm starting with five: across personal AI, smb platforms, commerce, native mobile, and public-sector transit. Each card says what was built, and what it took.

01fredispersonal AI advisor 02email hub9-agent email platform 03gerboniAI-assisted commerce 04ugokiAI-coached health app 05vtvtransit ops + AI agent
Honesty principle

Cards are framed as built / running / in-use, not as paid client engagements. Every claim is open to a source walkthrough in the build review. Overclaiming dies in the discovery call; I'd rather catch it before then.

01
fredis: personal AI advisor & second brain.
Shipping · in daily use
Flagship · 24-skill agentic stack
Source · private repo

What it is

A personal AI advisor with a 24-skill stack, persistent Obsidian-vault memory, and Slack / Gmail / Google Workspace / HubSpot integrations. Operates in advisor mode: every output goes to a draft for human review, nothing auto-sends. Built and running since early 2026.

Agentic capabilities

  • 24-skill stack with role-based personas, invoked contextually per request.
  • Advisor-mode safety boundary: drafts only, no send-authority on any external channel.
  • Persistent vault memory auto-loaded at session start.
  • Heartbeat / reflection / synthesis SDK loops for proactive scanning, memory consolidation, weekly synthesis.
  • Hooks-based safety enforcement: pretooluse blockers and session snapshots.

Stack

Python Claude agent SDK MCP / FastMCP tailscale serve https Obsidian markdown vault Hetzner VPS scheduled cron (heartbeat / reflection / synthesis)
Why it matters

The flagship proof: a working agentic advisor with persistent memory, real integrations, and a serious safety model. The exact shape of system I build for clients. I don't pitch agentic systems. I run on one.

02
email innovation hub: centralised email platform with 9 AI agents.
Built
9-agent platform · QA gates · RAG
Source · walkthrough on request

What it is

A centralised email development platform: build, preview, QA, and export production HTML emails from a single workspace. CMS-agnostic, security-first, GDPR-compliant. Designed for email teams of 5-50 developers across multiple esps. Built on personal time and stack.

Agentic capabilities

  • Nine specialised AI agents: scaffolder (brief → Maizzle HTML) · dark mode · content (subject lines / preheaders / ctas) · outlook fixer (mso + vml) · accessibility auditor (wcag aa) · personalisation (liquid / ampscript) · code reviewer · knowledge (RAG q&a) · innovation.
  • AI orchestrator routes between Claude and openai via a configurable provider abstraction.
  • RAG pipeline with pgvector embeddings and hybrid search.
  • Websocket real-time chat wires the AI agents into the editor surface.
  • 10-point QA gate: HTML validation, CSS support, file size, link validation, spam score, dark mode, accessibility, fallback handling, image optimisation, brand compliance.
  • Zero-trust API: jwt hs256, RBAC (admin / developer / viewer), row-level security on Postgres, aes-256 secret encryption, audit trails, brute-force lockout.

Stack

FastAPI (async) Next.js 16 react 19 Postgres + pgvector redis Maizzle (Node) tailwind / shadcn Docker compose alembic vertical slice arch
Why it matters

A production backend, modern frontend, and multi-agent orchestration in one product.

03
gerboni: AI-assisted e-commerce for Latvian heraldry.
Live · operating
Commerce · Pydantic AI · Stripe
Source · walkthrough on request

What it is

An online shop selling t-shirts featuring the coats of arms (ģerboņi) of major Latvian cities, with a 24/7 AI customer-support agent built into the storefront. Cultural-heritage product on a modern AI-native commerce stack.

Agentic capabilities

  • Pydantic AI agent powered by Claude sonnet 4: handles product discovery, order assistance, customer queries 24/7.
  • AI-driven product recommendations based on browsing history.
  • Admin analytics panel: sales, orders, products, users.

Stack

Next.js 16 react 19 typescript tailwind v4 (oklch) zustand radix UI FastAPI sqlalchemy 2.0 Postgres 16 Stripe Docker · k8s
Why it matters

A complete commerce build: real revenue, real payments, AI embedded as a product feature.

04
ugoki: AI-coached health & wellbeing app, iOS + Android.
MVP shipped · investor re-engaged
Native · AI coach · HealthKit
Source · walkthrough on request

What it is

A mobile wellness app combining intermittent fasting, HIIT workouts, and AI coaching for time-constrained professionals. A 15-20 minute daily routine for sustainable health optimisation. MVP shipped; 2026 re-engagement from walking ventures.

Agentic capabilities

  • AI coach with constitutional AI guardrails, cross-session memory, and skill-based prompt routing.
  • Bloodwork analysis: biomarker parsing from uploaded pdfs with trend tracking.
  • PubMed research hub with AI summaries relevant to user goals.
  • Health-device sync: apple HealthKit + Google Health Connect.
  • Gamification system wired into the AI's coaching context.

Stack

Python 3.12 FastAPI sqlalchemy 2.0 Pydantic 2.0 Pydantic AI Claude 3.5 sonnet / haiku Expo SDK 52 react native tamagui zustand tanstack query fly.io postgresql cloudflare r2 Expo push
Why it matters

Native iOS + Android, AI woven into the product rather than bolted on, real health-data integrations.

05
vtv: unified CMS & AI agent service for Riga city-bus operations.
Built · presenting to Rīgas Satiksme
Civic · GTFS · bilingual lv / en
Source · walkthrough on request

What it is

A FastAPI + Pydantic AI multi-agent service paired with a Next.js 16 / Turborepo CMS frontend for Riga's municipal bus system. GTFS-compliant, swappable LLM provider, bilingual (Latvian + English), 4-role RBAC. Presentation secured with Rīgas Satiksme representatives; source repo to be made public after that meeting.

Agentic capabilities

  • Four specialised agents over one shared tool registry: a domain router picks transit, knowledge, vault, or a unified fallback per request.
  • Five read-only transit tools: advise, don't decide.
  • Four Obsidian-vault tools with dry-run previews.
  • Swappable LLM provider via single env var: anthropic, ollama, openai, groq, openrouter.
  • Openai-compatible streaming chat endpoint.

Stack

Python 3.12 FastAPI 0.120 Pydantic AI sqlalchemy 2.0 async postgresql 18 alembic mypy + pyright (strict) ruff pytest uv Next.js 16 react 19 tailwind v4 shadcn/UI auth.js v5 next-intl Turborepo + pnpm 75 tests < 1.2s
Why it matters

Public-sector scale with regulated-sector requirements: GTFS, RBAC, safety guardrails, bilingual delivery.

Seen one that looks like the system you'd want built?

30 minutes. One workflow. Honest answer either way.

Book a build review