In the places where the work really matters, AI stays out. The model is good enough. But afterwards nobody can show what it did or who agreed to it, so the answer stays no. I change that two ways: an engagement that starts with a map of your system and ends in evidence you can re-run, or your own AI on Spegling.
Done for you · engagements
The person who holds the code is leaving. AI made a change and something broke quietly. Someone is coming to look at your codebase, and you would like to look first. A reading gives you the map in a week, from €1,000: what depends on what, who holds it, where a change breaks something quietly. Building and proving come after, sized by what the map shows.
See the engagements →Run it yourself · Spegling
For anyone whose AI life has become chaos: many assistants, many accounts, every one starting from zero, work and private tangled together. Spegling is the part that stays. One memory every AI you use shares, real walls between your lives, work checked by a second AI from a different company, and a record you own and can take with you.
Open Spegling →Both run on the same machinery. Past the map, every change lands with its evidence, and the record starts on day one.
Varjosoft is one person, deliberately small. Before you weigh any claim below, read something and see how I think. The products can wait; they are built on the same habits as the writing.
Every field that carries real consequence learned it long ago: when it matters, you keep the record as it happens, because the account written afterward is worth a fraction of the one written in the moment.
The highest-leverage thing to build is not the model. It is the record the model runs on, captured as it happens, because that is the one part you cannot add later.
One AI does the work. A second one, built by a different company so it is not grading its own homework, answers a fixed list of yes-or-no questions about it. One verdict per question, no vibe-check. What passes is filed to Memory and sealed onto the Chain with its verdicts, its cost, and the hash of the row before it.
The stop is proportionate: it closes only on what you cannot take back, and lets reversible work run free. The record, though, is complete. Holds included.
Your memory, your review, your record. The speed of AI, without losing track of what it is doing.
Spegling
The part of your AI that stays while the models change under it.
One memory across every session and every agent you use: pour into the Journal for three minutes, or wire your Claude or ChatGPT in over MCP, and tomorrow already knows what today was about. Work and private stay in separate rooms, and the boundary is a column in the database, not a promise in a prompt. Important work is checked by a reviewer from a different family before it files, an unsettled question routes to you instead of becoming a guess, and everything seals to a record you can export and keep. Longer sessions run the boring way, which is what lets you close the laptop.
spegl.ingDots
AI that earns the right to answer. A widget for any page with words worth reading: after two minutes of real attention it answers questions grounded in what is actually written. Refuses off-topic. No cookies, no tracking.
Live in the corner of this page. Spend a few minutes reading and it will offer to answer.
patterns-starter
The open bootstrap for your own way of thinking. One Markdown file: fork it tonight, add your first pattern this week. If Spegling stops existing tomorrow, your patterns still work.
Fork on GitHubMemory, decisions, and evidence are owned. Inference, calendars, and email are commodities. That selective sovereignty is what keeps you less capturable as the tools keep changing underneath you.
From reading the system you have to proving the AI you run: independent evidence, and the boundaries that turn "not allowed" into "allowed, with receipts." Five shapes at a fixed scope and a fixed price, or embedded part-time with your team when the work runs longer than a hand-over. One person either way. Or just a conversation, if you are not sure where to start.
Read
Your codebase mapped before AI touches it. Who holds it, what breaks quietly.
Adopt
Find where AI fits, with a roadmap and the boundaries drawn.
Prove yours
Independent evidence on an AI system you already run.
Build
The AI for one job, built and proven. Not a demo.
Keep it honest
Kept current as the models change underneath you.
Also on the deep-technical track: quantize, serve, prove it. Third-party validation of large-model serving on vLLM: real hardware, reproducible numbers, publishable artifacts.
Hannu Varjoranta. Systems engineer, founder, writer. Two decades of infrastructure, security, and data systems, including Spotify and F-Secure; co-founder of Valo and Cloop. My working days belong to Avrea, where I’m a founding engineer on its production platform. Varjosoft is what I build around that, on my own time.
Varjosoft is intentionally small, personal, and long-term. The work here is the layer above execution: memory, trust, and the governance of systems that act on our behalf.
The path is a conversation. Spegling access is by invitation: say what you would run through it. Engagements are open, scoped to fit around a full-time role: say what the system must do and what it must never get wrong.
Or write directly: hannu@varjosoft.com. A person reads it and a person answers.