Hallengren’s · Systems

The systems, and who they were built for.

Not a shop. A map of what has been built and what it runs. Some are systems under the Hallengren’s flag, each at its own stage. Others are whole products built for someone who had a need, now living lives of their own. This page is the way in; each system has its own address.

Under the Hallengren’s flag

Three systems, three stages. All mine, end to end.

These carry my name because I own every layer of them, from the strategy down to the server. Each sits at a different point in its life: one shipped, one in beta, one still in the lab. That range is the point. It is what building looks like when it is a practice, not a launch.

Hallengren’s Pillar Two · Live

The system that solves the problems around Pillar Two.

The OECD's global minimum tax turns a multinational's position into a number that has to be exactly right, traceable to the rule it came from, and defensible to an auditor. Most of the market answers with spreadsheets and consultants. This answers with a system: a deterministic engine where the maths must be exact, the model confined to one job and never let near the calculation.

It is built to distrust its own reading. Every figure is extracted three times, by three genuinely different techniques — one of them never a model — and the readings are checked against each other; disagreement is the signal. Then a human signs off on every document before anything is calculated, each number highlighted on the very page it came from. The system refuses to approve what it cannot vouch for.

the auditor, two years later“Walk me through this effective tax rate.”

what the filing answers“Every number carries its OECD article — the trail runs from this figure, through each adjustment, back to the page of the annual report it started on.”

That is what defensible means: not an opinion — a trace.

Where the rules leave a judgment call, your own tax people take the position — on a slider from conservative to assertive, within the room the law itself leaves — argued by three perspectives on the same question: the tax counsel, the consultant, the CFO — drawn from a knowledge base spanning 58 jurisdictions — 325 grey-zone cells and 57 transfer-pricing profiles, every claim adversarially reviewed against primary sources. Every move recomputes the top-up for real. The same discipline now reads your intra-group flows: arm’s length lit up from both countries’ reviewed rules at once, while only the country pair and transaction type ever cross the wire — never your amounts. And your positions never leave your machine: they live in a local file with you, not with us.

deterministic · OECD-traceable · defensible · multi-jurisdiction · in production

Annual reports · PDF
an AI reads code verifies a third arbitrates
three genuinely different techniques — disagreement is the signal
A human signs off on every document approval stays locked while flags remain
pure calculation from here down
GloBE engine · per jurisdiction
1Financial net income
2GloBE Income
3Covered Taxes
4ETR & Top-up if below 15% — QDMTT or IIR?
conservative recommended assertive
every position recomputed for real — argued by tax counsel, consultant and CFO lenses
Intra-group flows · arm’s length both countries’ reviewed TP profiles — only country pair + type cross the wire
Audit invariants every number carries its OECD article
GIR-XML — validated against OECD’s official schema → your positions and intra-group amounts never leave your machine — a local file with you, not with us

The AI never touches a number — and no number goes unverified.

Human-in-the-loop by design: the machine does the heavy reading and shows exactly where it is unsure — the judgment stays with your own tax people. GIR-XML is the return every in-scope group files; here it leaves with the audit trail built in.

Hallengren’s Yggdrasil · Beta

The world-tree for people who orchestrate agents.

If your day is a fleet of AI agents across a dozen projects, the tools are scattered: a terminal here, a voice app there, a memory that never talks to either. Yggdrasil is one suite where each part has its place. The myth is the architecture: Odin decides in Asgård, Hugin carries the thought out, Munin returns with the memory, Heimdall watches the bridge. One system, in the shape of a story a thousand years old.

It runs on your machine, and the agents are your own — Yggdrasil’s job is judgment, not runtime. Odin is what makes it a system rather than a dashboard: he weighs every task against your goals, for the project and for your life, and decides what deserves your attention next. Mimer holds your playbooks — the way you already work, written down — so no agent ever starts from a blank prompt.

your agents your memory your keys your Mac — everything lives here your phone bifrost commands in · results back local by default · sync opt-in

you, from your phone“The tests on main are red — fix them before standup.”

what comes back — twenty minutes later“A green diff, waiting in your pocket. Odin read the repo’s memory, chose your debug playbook from Mimer, and an agent ran it in Heimdall — while you made coffee.”

Command in → judged → run → result back. The laptop never opened.

one app, four rooms · voice → priority → terminal → memory · in private beta

Asgårdthe app · four rooms · runs on your machine
2 the orchestrator Odin reads your goals, memory and agent state — decides what runs next
Huginthe thought · voice in
Muninthe memory
Mimerthe playbooks
3 Heimdallwhere your agents work · the default view
agentagentagentagentagent
Bifrost · desktop statusbar companion
1Bifrost · iOS commands from anywhere4
1command in2Odin decides3runs in Heimdall4result back

Your agents, your machine, your keys — built for how fast you can verify, not how much it can automate.

The agents are Claude Code, Codex or Gemini CLI in real terminals — auto-resumed if the machine goes down. Munin is a local Postgres every agent shares over MCP. Nothing leaves your Mac unless you turn sync on.

Hallengren’s Symbios · R&D

The hardest problem I know: relationships, taken clinically.

Symbios started from a simple thesis: all therapy is relational work. It is a research effort toward a genuinely clinical tool for couples therapy, relationship work, and therapy more broadly, with an agent that holds each person's side. I keep it here, openly marked as in the lab, precisely because of its difficulty. The complexity is the reason it is interesting, and the reason it takes the time it takes.

The part people fear is the part it is built hardest against. Nothing is shared unless you choose it: you decide what may cross, and at which of three levels of transparency — per relationship, and you can change it whenever you like. It cannot be milked, either: there is no way to ask for your partner’s worst moments, no report to pull, nothing to use as ammunition — it only ever carries across what serves the relationship.

you, to your guardian“She’s been distant for weeks. I’m starting to think she doesn’t want this anymore.”

what comes back — at the level she chose to share“I’m sorry it feels that way — it isn’t about you. She’s carrying something she isn’t ready to share with you yet. Right now, the strongest thing you can do is be her support.”

That is what sharing looks like. Never a transcript — only what helps.

And nothing accumulates against you. Your conversations leave no journal to reread — as with a clinician, the record is not yours to carry — and every session starts clean. What endures, anonymised behind the filter, are the patterns that help it understand you: never your words from a bad night. And what it holds about you — you personally, never anyone else — is yours in full: see it, correct it, erase it, whenever you decide.

agent-per-person · relational · clinical ambition · under active research

Two people — any relationship — each talking freely with their own guardian AI. Everything crosses an anonymisation layer before the clinical stack — an agent standing on layers of clinical knowledge and anonymised memory; recommendations come straight back. you you talk freely your guardian loyal to you alone any relationship — one sealed world each your partner their guardian loyal to them only what you chose to share the anonymisation filter names & identities stripped here the clinical agent clinical knowledge continuously updated anonymised memory all your relationships & emotional patterns your goals & agreements what you work toward · what was actually agreed the clinical stack advice weighed against all of it — not a bare LLM
you ⇄ your guardian an AI loyal to you alone your partner — or a parent, a friend ⇄ their guardian any relationship · one sealed world each ↓ only what you chose to share, at the level you set ↓ into the anonymisation filter — names & identities stripped, nothing passes raw the clinical stack clinical knowledge, continuously updated · anonymised memory of your relationships & patterns · your goals & agreements — advice weighed against all of it, not a bare LLM recommendations go straight back to your guardian

The other person never sees what you said — only what you both agreed to share.

Every relationship is its own sealed world — what you share in one never reaches another. The stack never learns who anyone is; its recommendations go straight back to your guardian.

Built on commission · proof of concept

Someone had a need. I built the system. It took on a life of its own.

These two are not under my flag; they are their own companies, their own brands. But they came from the same place a commission does: someone with a need, and one person to turn it into a working system, end to end. No one paid me a fee for these. They are simply proof that the model works, running in the world today.

Live · a company now

Autoply

The commission: David Bergqvist had the idea. I developed it with him, productised it, and came in as a partner. What began as a concept is now a company with its own product.

Live · its own entity

Vuelette

The commission: Michelle had a need — a better way for photographers and their subjects to work closely together. I built the system to serve it. It now runs as its own business.

One hand, one doctrine

Different problems. Different worlds. One hand.

A tax engine, an agent suite, a clinical research system, a company built from a founder's idea, a tool built from a wife's need. They share no code and no market. What they share is that one person carried each from the first thought to the running system. If you have a system that does not exist yet, that is the whole offer.

Pillar TwoYggdrasilSymbiosAutoplyVuelette One hand · one doctrine