ESALI · STOCKHOLM · SINCE 2014TWELVE YEARS
esali is a Stockholm venture practice. It starts with blue-sky thinking, looks for where the world is missing something, and builds sovereign technology to close the gap. Personal AI, virtual worlds, property intelligence. Independent since 2014, answerable to nothing but whether the idea actually helps, and keeps working.
The record, newest first. Every number below is real and dated.
Property intelligence for Sweden. Paste any home listing and see what the public record knows: the housing association’s real finances, the schools, crime, environment, flood and ground risk, before you buy.
Your life’s OS, with a life of its own. You can download an AI. You can’t download the one that has known you for years. V holds your whole life, learns you deeper every day, and rebuilds itself while you sleep, so it grows more yours, and more irreplaceable, the longer you keep it. The model is swappable; the years of you inside it are not. We don’t ask you to imagine what we can build. We point at V.
Builds immersive Minecraft worlds that have reached millions of players. Worlds you’ll want to stay in, made to keep working long after launch.
esali is not an agency. The ideas start inside the practice, and the practice carries each one all the way, until it runs in the world, and keeps running.
Not decks, not pilots that quietly die. Software in production, holding real data, doing real work.
The people who plan the work are the people who do it. No account layer, no hand-offs.
esali began in 2014 doing what clients needed, consulting and campaigns for other companies. Useful work, but never the point. What it does now is build products it keeps and runs itself: Plank, immersive worlds that have reached millions of players; V, a personal AI you own outright; Plum, property intelligence for Sweden. Sovereign things, answerable to nothing but whether they actually work, and keep working.
The three above are what you can use today. Under them sits a deeper stack esali builds for its own products, most of it designed as separate layers that can be lifted out and stood up on their own. A sample below, tagged honestly: what runs today, what is building, what is early.
A private, continuous memory that persists across every session, on hardware you own rather than a vendor's cloud. It is layered: permanent directives that never fade, plus episodic recall, all scored by how surprising or important a moment was so the signal rises and the noise decays. It runs on local embeddings and keyword search fused together, and it links related memories over time, so it recalls not just facts but the connections between them. The core of the system, and the hardest part to copy. The one thing you can't download: it's owned, on your hardware, and compounds, so the longer you keep it the more irreplaceable it gets, and it can't be lifted out in a migration script.
The system watches its own behaviour, finds the patterns that reliably work, and promotes them into standing rules it then enforces on itself, with no language model in the loop. A pattern has to earn it: enough confidence and enough repetitions before it hardens, on a rolling cycle, so habits form the way they do in a living thing rather than being hand-coded. It governs by argument too, with separate roles that challenge and can veto an action rather than one model rubber-stamping itself. Governance that legislates itself with no AI model in the loop: deterministic, auditable, and free to run. Real autopoiesis, not a prompt.
Give it a goal like "make this better" and it builds the crew to do it: a panel of expert agents and a spectrum of real user-personas, each with a distinct role and the right model chosen for that role. It sets how they work together and the flow they run, has them test and rebuild against the live thing, scores and ranks every proposal with its own rating, loops until the gains flatten, and brings you the decisions as one-tap approvals. It routes, yes, but it is an unusually intelligent router with far more breadth than routing: really a chief of staff. Dynamic agent-team assembly plus governance: it composes the right crew, roles and models for a goal and governs how they argue to an answer. Nobody else assembles the sovereign, governed whole.
The default posture is defensive: it refuses to act on state it has not verified, and throttles itself rather than guess, where most AI defaults optimistic. A supreme stop authority can halt everything, an ethics layer can veto an action outright, and a set of deterministic checks watch the machine itself, disk, ports, processes, on a tight loop with no model involved. Secrets live in a separate vault, never in the reasoning. If something is uncertain, it stops instead of pressing on. AI that fails closed, refusing to act on what it cannot verify, where every other assistant defaults optimistic.
Nothing counts as done without proof: work is classified as running, partial, documented-only or broken against real evidence, not against what the AI says about itself. A separate conscience runs on a different AI vendor on its own schedule and asks the blunt question, is the system lying to itself, and it has caught real problems that the builder missed. A drift check re-runs regularly so a claim that was once true cannot quietly rot. A second conscience, on a different AI vendor, that catches the system claiming done without proof or believing its own spin.
Scout is the System for Contextual Observation and Update Tracking, the external half of the system's situational awareness. It is not a feed reader: it proactively scans the AI frontier, arXiv, Hacker News, Hugging Face, Reddit, YouTube and email, compares what it finds to the system itself, synthesises across outlets so the same story from many sources becomes consensus-versus-disagreement rather than noise, and judges what is worth learning, contributing or defending. When it spots a model materially better than the current one, that feeds an automatic benchmark before adoption. The scanning runs live and current today; a conversational research surface, think Perplexity, but yours and sovereign, is the planned front end. A research analyst that watches your whole field for you, takes a position on what matters, and feeds it straight into a system that can act on it, all on hardware you own.
HumSim, human simulation testing, drives real, signed-in browser journeys as a spectrum of simulated user-personas, first-timer, family, investor, downsizer, skeptic, price-anxious, time-poor and more, against your live product, so you see where real people actually hesitate and bounce, not synthetic lab metrics. It is the test-group engine the orchestrator uses to improve a site, and it is getting more advanced, voice-driven journeys, wider persona range. It tests your product the way a crowd of real, different humans would, on the live thing, and hands you exactly where they got stuck.
The most developed piece is a peer-set forecaster: for a new launch it reads the history of comparable ones, groups by theme, and returns a likely range and a midpoint rather than a single false-precise number, weighted by how many real comparables it has and by recent momentum, and it flags plainly when the sample is thin. It predicts, then checks the prediction against what actually happened, so its confidence is earned, not asserted. Today these forecasters live in separate corners; the work in progress is one shared prediction store so every surface writes to it and calibration is computed once. Forecasts that score themselves against what actually happened: a calibrated range instead of a confident guess, and it tells you when the data is too thin to trust. A tool for predicting from data, and the seed of how the system models the world.
It drives a real browser signed in as you, so it reads what sits behind a login, past bot walls and identity checks a plain scraper never gets through. Today this is a sensor: it sees your actual accounts and pulls the real state, alongside native connections into email, drive, calendar and your task and contact tools. Acting for you inside those sessions, filling and submitting on your behalf, is still in design, not yet switched on. It works behind your logins, as you, reading and (soon) acting on what a scraper or a public API never reaches.
One place for every channel. The design: email, Slack, Discord, WhatsApp and social all route into a single inbox, an assistant triages what came in and drafts how to answer each, and you decide, send, or let it auto-reply on rules you set. The surface exists (reach.esali.com), but the honest state is that the unified pipe carrying every channel into one plane is not fully wired yet, so this is being built, not shipped. Every channel in one place with an assistant that drafts the reply and governance you control, running on your terms instead of inside someone else's silo.
It reaches out on its own rhythm, not only when you ask: a morning brief, a midday check, an evening digest, plus nudges through the day and an urgent alert that cannot be silenced when something genuinely needs you. It reconciles your open life fronts on a schedule, reading your own notes, tasks and calendar, so the brief is about your real situation, not a generic feed. The timers are still restart-fragile, which is why this is building rather than rock-solid. Most assistants wait to be asked; this one runs on its own clock and speaks up first, about your real situation, not a feed.
A bookkeeping and accountant-collaboration portal, in daily use for esali's own accounts: it works through the accounting cycle, resolves gaps, links each transaction to its evidence, produces the VAT summary, and shares a locked-down quarter view with the outside accountant. The running cockpit is single-company today; a wider collaborator portal for a client's own advisor is staged and not yet switched on. A real product for small businesses: it cuts the cost and grind of bookkeeping and gives your accountant a clean, locked-down quarter to work from.
A browser extension that exports your AI conversations out of ChatGPT, including the Teams workspaces nothing else can export, into open formats (JSON, Markdown, CSV) in minutes. Custom GPTs, projects, canvas, memory and instructions come with them. It runs entirely on your machine, so your data never leaves it. Open source, MIT. The only tool that exports ChatGPT Teams workspaces, client-side and open.
A Swedish renovation-proposal tool for contractors, builders, electricians, plumbers, that turns a job into a scoped offer to hand a client. It exists as an assistant and an early site rather than a finished product, and it is set aside for now, not gone.
Each of these is a layer of the same system. Because they are built as layers, any one can be stood up on its own. If one fits a gap you have, that is a conversation.
If you have a system that needs to actually work, we’d like to hear about it. No pitch, no funnel, just a conversation.
Email esali Work with esali on LinkedIn →