saulera · the work, presented

one founder. a team of agents. five solutions.

I build software with agents at a pace that used to need a team, without skipping the gates a good engineering team would insist on: tests first, strict types, injection defence, and the legal lines checked before anything ships. This page shows the five solutions that prove it: what each one is, how it was built, on what stack, and what guards it. Every one is open to a live walkthrough, code and all.

01

fredis

shipping · in daily use

the second brain the company runs on

what it is

fredis is a personal AI advisor: a self-hosted second brain that reads my inbox, calendar, CRM and code, drafts everything, and sends nothing. Twenty-four advisor skills carry the playbooks: a CTO, a CISO, a product manager, a launch governor, a meeting scribe. Memory lives as plain markdown in an Obsidian vault, versioned in git and readable without the AI. If the model goes away tomorrow, the brain stays.

A heartbeat ticks through the day. It scans the vault, checks every project against its go/no-go gates, drafts what is due, and posts one line to Slack: here's what I'd do if I were you. You decide.

the team

CTO
takes a repo issue or a client brief and returns a draft pull request. It works out of the studio below and never touches the vault.
PA
triages inbox, calendar and admin. Reads everything, drafts replies, sends nothing.
sales
scoping-call follow-ups and CRM pipeline nudges. Draft-only by design; cold outreach never auto-sends.
marketing
drafts and repurposes content into a review queue. The lowest-risk chair on the team: pure drafts.

The CEO chair stays human. Every send key stays with me. Zero unauthorised sends since day one.

the studio

Build work runs through a harness with a cockpit UI. Every job gets an isolated git worktree, a run of the checks, and a route back as a draft pull request; the cockpit shows every run, every draft and every decision waiting on a human. That is how fredis builds applications and digital products, and how everything on this page is maintained.

how it was built

Python 3.12 on the Claude Agent SDK. Hybrid memory search, vector plus keyword, over local embeddings: SQLite at home, Postgres with pgvector on the server. It exposes its own MCP server for remote access. Hooks guard the boundaries: secrets blocked before they leave, the SOUL file (its constitution) read-only, prompt injection screened twice, once by pattern and once by a model. It runs as systemd loops on a Hetzner VPS.

the numbers

24advisor skills
741tests
12+live integrations
0unauthorised sends

the stack

02

email innovation hub

live mvp

production email, built and checked by agents

what it is

Email development is stuck: five to eight disconnected tools per team, and a rendering war with more than ninety email clients. The hub replaces that with one workspace. A code editor, a live preview and an AI chat sit side by side. Nine specialised agents do the repetitive work: a scaffolder that turns a brief into working HTML, a content writer, a dark-mode fixer, an Outlook fixer, an accessibility auditor, a personalisation agent, a code reviewer, a knowledge agent that answers from the documentation, and an innovation agent that proposes what to try next.

A self-correcting pipeline runs their output against a ten-point QA gate: rendering, file size, spam score, accessibility, brand. Work that passes exports straight to the ESP, and a client portal handles approval with an audit trail.

how it was built

Built solo with Claude Code across fifty planned phases, then a formal seventy-finding tech-debt audit. Vertical slice architecture with strict MyPy and Pyright, 27 MCP tools, and a Node sidecar compiling Maizzle on every save. Seven Docker services behind Nginx.

the numbers

9AI agents
261API endpoints
7,055tests
867commits

the stack

03

vtv

presented at riga city hall

transit operations for Riga's municipal buses

signal

Presented at Riga City Hall to Edvards Šlesers and his colleagues in late May. City Hall is forwarding the documentation to Rīgas Satiksme, Riga's municipal transport operator.

what it is

vtv is a unified operations platform for public transit, starting with Riga's buses and built to extend across Latvia. Live vehicle tracking from GTFS feeds, proximity search on PostGIS, a bilingual CMS in Latvian and English with four-role access control, and an AI agent in the sidebar that answers from live operational data.

The agent's transit tools are read-only on purpose: it advises, people decide. The model is swappable across five providers with one environment variable, so a municipality is never locked to a vendor.

how it was built

FastAPI and Pydantic AI on the backend, vertical slices, strict typing. Redis caches the live feeds and WebSockets push them to the screen. The CMS is a Next.js 16 Turborepo monorepo with a design-token system and a TypeScript SDK generated from the API. Twenty-one project-specific Claude Code commands drove the build: prime, plan, execute, validate, commit.

the facts

GTFS-RTlive tracking
lv · enbilingual
4access roles
5LLM providers

the stack

04

ugoki

mvp complete · investor re-engaged

an AI-coached wellness app for iOS and Android

what it is

ugoki fits health into the day you actually have: intermittent fasting, workouts of fifteen to twenty minutes, and an AI coach that remembers you between sessions. The coach runs on constitutional guardrails covering safety, evidence and personalisation, and reads context from Apple Health and Google Health Connect. Upload a blood test and it parses the biomarkers and tracks the trend. A research hub summarises the PubMed papers behind the advice.

how it was built

Eleven independent backend modules that only talk through defined interfaces; no module touches another's tables. The mobile app is Expo and React Native with a shared design system. Built docs-first with Claude Code and seven project-specific skills. The shipped MVP brought the original investor, Walking Ventures, back to the table this year.

the numbers

61klines of code
183API endpoints
11backend modules
114exercises

the stack

05

gerboni

live · taking payments

a live shop with an agent on support

what it is

gerboni sells t-shirts of Latvian city coats of arms. Stripe payments are live. A support agent sits in the storefront around the clock, answers questions and recommends products; an admin panel tracks the numbers. It is the smallest system here and the most honest one: a real shop, real payments, an agent doing a real job.

how it was built

Next.js 16 storefront, FastAPI backend, a Pydantic AI agent on Claude for support and recommendations. Deployed with Docker; live and operating.

the stack

the method

delivery that scales without headcount

One operator, one method. Every build starts with a lean PRD and runs through agent workflows in Claude Code: prime the context, plan the feature, execute, validate, commit. When a system grows it gets a formal audit; the email hub's audit covered seventy findings, and vtv and fredis run the same discipline.

the gates every build clears

tests first
7,055 tests on the email hub, 741 on fredis, mutation and property tests where it counts. Red means it does not merge.
strict types
MyPy and Pyright in strict mode on every Python backend; TypeScript SDKs generated from the API so the frontend cannot drift.
vertical slices
the email hub and vtv are built as vertical slices: every feature owns its own routes, service, repository and models. ugoki's eleven modules talk only through interfaces. A slice can be audited, rebuilt or rolled back without touching the rest, which is exactly the shape agents build best in.
prompt injection
screened twice on fredis, once by pattern and once by a model, with trust boundaries around anything external; a scanner guards the email hub.
agent guardrails
the ugoki coach runs on constitutional rules, safety and evidence before helpfulness, with a judge model scoring its answers. Every email the hub's agents produce must clear the ten-point QA gate before it exports. On fredis, hooks block dangerous commands and secrets before they reach a shell, and its core instructions are read-only to the agent itself.
data security
Postgres row-level security for tenant isolation, AES-256 on stored credentials, short-lived JWTs with revocation, role-based access on every multi-user system.
agent containment
outbound channels write-locked (fredis sends nothing), transit tools read-only (vtv advises), deletes behind confirmation and dry runs, file paths sandboxed against traversal, spending caps on the model APIs.
the legal lines
an IP guard that fires when product work touches employer time, draft-only outreach because cold email is regulated, GDPR in mind from the schema up.

The point of all five systems is the same bet: the cost of build number n falls as the workflows compound, and the gates travel with them. That is the company.

the product

your own advisor, under a name you choose

Everything above compounds into one product. fredis and the build harness are a single system now: the same agents that run my inbox, pipeline and projects also build applications and digital products, watched from a cockpit that shows what is building, what is drafted, and what waits on you.

That system is what saulera sells to small and medium businesses: a personal advisor and assistant, set up in a simple web UI. fredis is what I named mine. You name yours in the wizard, and it answers to that.

1

open the URL

As a saulera client you open a link in the browser. No install, nothing on your machines.

2

run the setup wizard

Name your advisor, answer the intake questions, choose the chairs on your team: PA, sales, marketing, a CTO that builds.

3

connect your services

Gmail, calendar, CRM, files. The tools you already run on, connected read-only until you decide otherwise.

Then it works the way it works for me: it reads your business, drafts everything, and sends nothing without you.