Expertise, encoded
The hard part is rarely the software. It is getting a judgement that took twenty years to form into a form a machine can execute repeatably — and being honest about the parts that did not survive the translation.
We build the systems that expertise turns into.
Evander Labs is a UK software practice. We take domain knowledge that normally lives in one person's head — how a complex system actually behaves, how a record actually gets made — and we build it into software that runs, is tested, and leaves an audit trail behind it.
What we do
Not an agency. Not a consultancy. An engineering house.
Every product here started as a problem someone actually had, and every one of them is running code rather than a deck. Three things hold across all of them.
Evander Labs sells software and software / web development services. Our data and analytics products are tools for research, simulation and education. We are a software practice, not a financial services firm: we provide no regulated financial services or advice, and we never hold, transmit or handle client money.
Institutional standards, small team
Version control, tests, documented interfaces, reproducible builds and written evidence for every claim. The size of the team is not a reason to lower the bar; it is a reason the bar has to be automated.
Guardrails designed in
Limits, kill-switches and fail-safe defaults are part of the architecture, not a later release. A system that cannot be stopped cleanly is not finished, and a system with no cap on what it can spend or send is not safe.
Portfolio
Thirteen products across three lines.
Three lines of work, each with its own constraints. What follows is what each product does, in plain terms.
Data & analytics
05EvanderOS
The flagship: a data and analysis operating system. A kernel, FORGE, exposes four sockets — data, strategy, test and critic — and every tool is a ring-fenced sub-application running against them.
Research, simulation and education only
pipSave
A research lab for sequential decision-making under uncertainty. A beam-search planning engine with a champion/challenger learning loop, trained on over a million mined historical episodes, keeping a full audit trail of every decision it makes.
Research environment — simulation only
PipTax
A content and search engine covering cost analysis and comparison — what a thing is advertised to cost against what it effectively costs, computed from the underlying data.
Golden Risk Index
A published news-sentiment index for gold, derived from news across sixteen languages through fourteen transmission channels and resolved to a single daily reading.
A data measurement — not a recommendation
Proving Ground
An independent cost-audit and comparison platform: what a given account actually costs to run, computed rather than asserted.
Creative AI
03Tygeta
An owned end-to-end music generation engine. Not an API wrapper: the full stack, including a 288,000-note MIDI corpus and a JUCE-based audio effects chain we wrote ourselves.
EvanderFlix
Music-video generation: a track in, a finished cut out, with a consistent visual identity held across every shot rather than re-rolled each time.
12second
Short-form marketing video for businesses, built on the EvanderFlix engine and tuned for the twelve-second slot a product actually gets.
Platform & tools
05Orchestradi
A unified API gateway with a compliance-first architecture: one interface over many providers, with the permission boundary enforced at the gateway rather than trusted to each caller.
Xevander
End-to-end encrypted messaging built on Matrix and Synapse, self-hosted and administratively provisioned, for organisations that need their messages to stay theirs.
Valo
A messenger-first assistant that works where the conversation already happens, with an explicit capability matrix governing what it may do unattended.
ArriviGo
A concierge that books things — travel, tables, tickets — from a plain request, with hard rails around spending and outbound communication.
RIO
A client portal and reporting platform: statements, activity and periodic reporting generated from source records rather than assembled by hand.
The pipeline
What is on the bench now.
Work moves in one direction: research establishes whether something holds, development makes it usable, release puts it in front of people. Nothing skips a stage.
Now
EvanderOS kernel and the first ring-fenced tools
Hardening the FORGE sockets and moving the existing analysis tools onto them, so each one runs inside its own boundary with its own data contract instead of sharing a process.
Now
Cost modelling as a first-class primitive
Costs, latency and friction modelled explicitly everywhere a simulation runs. Most published research quietly assumes these away; when you put them back, a great many results stop being results.
Next
Tygeta from engine to product
The generation stack works. The work now is the surface around it: a render queue that survives a failed node, and an editing interface that treats output as material rather than a finished artefact.
Next
Orchestradi as shared substrate
Moving every product's outbound provider traffic behind one gateway, so rate limits, spend caps and audit logging are configured once rather than re-implemented per application.
Later
Self-hosted inference
Bringing the language and audio models that sit under the products onto our own hardware, for cost, latency and — mostly — to stop a roadmap depending on somebody else's deprecation schedule.
How we build
The evidence is part of the product.
Anyone can claim a system works. These are the specific mechanisms that let us show it — and let us find out quickly when it does not.
Hash-chained evidence trails
Decisions and results are written to append-only ledgers where each entry carries the hash of the one before it. A record cannot be quietly revised after the fact — the chain breaks, and the break is visible.
Fail-safe defaults
Every automated component has a latch it can fall into and stay in. If a data feed goes quiet, the dependent system pauses rather than acting on stale information, and it does not resume on its own.
Simulation before anything else
Analytical tooling is built and validated against historical and simulated conditions, with realistic costs applied. That is the boundary of what these products do.
Out-of-sample discipline
Findings are held to data the work has never seen. A result that only exists in the period it was discovered in is treated as a null result, and recorded as one.
Documentation in the same commit
Code and its documentation land together — DOCS.md is not a
separate task that never happens. Every commit records what changed, why, and
the state it left the system in.
Spend and capability caps
Anything that can call a paid service has a hard daily ceiling, and anything that can act on someone's behalf works from an explicit allowlist. Both fail closed.
Contact
If any of this is useful to you, write.
We are a small team and we read everything. Technical detail is welcome and shortens the conversation considerably.
Business
Evander Labs
Scotland, United Kingdom
What helps
Tell us which product this is about and what you are trying to do. If it is technical, the detail is the useful part — send it.