Custom software and our own AI product
Software that can show you its working
We build, to order, the software a firm is currently running by hand — and everything we ship can tell you where its numbers came from. On this site that is not a promise: the examples below run the real logic in your browser, with every rule on show.
- You own the code, always
- A plan you keep either way
- No case studies we invented
Systems you already run
Connect
One integration layer, monitored
AI + Automation
Extract · classify · route · act
Your workflow
Unchanged where it already works
Measurable outcome
Hours returned, errors down, status visible
Illustrative flow — Twishiv Platforms builds the middle two layers.
The problem
You already have the software. The problem is everything happening between it.
Two of these from each of the three industries we work in. None of it appears in a process document, and all of it costs hours every week.
The deck goes out and nobody can swear the numbers are right
Three people rebuilt it from four spreadsheets, and the figure that gets quoted back at you in a meeting came from a file nobody can name.
Scope creeps and nobody notices until the invoice
The change was reasonable, the engagement letter says something else, and the conversation happens after the work is already done.
Month-end is three people retyping PDFs
Vendor, date and every line item, keyed by hand, with a transcription error rate nobody measures and a duplicate caught at the reconciliation.
Two systems that are each fine alone
Together they generate a reconciliation job that arrives every morning, and a habit of entering the same record twice.
Screening sixty CVs in an afternoon
A skim for keywords, under time pressure, that quietly rejects people who could do the job and simply wrote less.
The spreadsheet that runs everything
Business-critical, understood by one person, one bad paste from a very long day.
Who we build for
Three industries we understand before the first meeting
We are small, so we would rather be genuinely useful in a few places than plausible everywhere. Each of these pages links something you can run.
Consulting
The deliverable is the product. We build the production line for it — and we already built one for ourselves.
What we build for them
Finance operations
Month-end should not depend on someone retyping PDFs. We build the matching and the exception flow around the system you already run.
What we build for them
Staffing
Software for the firms that place people, built around the ATS you already run — and a company that will never compete with you for a placement.
What we build for them
Not one of those three? We will still take the work on — scoped the same way and built to the same standard. We would simply rather tell you where we are strongest first.
Not a demo video
This is what one of the examples actually returns
Rendered by the same engine the example runs on, from the same synthetic data. Change a rule and this panel changes with it.
Half of what this candidate is “missing” was never in the CV to begin with. Treating that as a gap is how screening software rejects people who can do the job and simply wrote less — and not being able to tell the difference is what we mean by software that cannot show its working.
The reasoning behind this panel is written up in why keyword screening rejects good candidates, one of the reference guides written to be useful whether or not you ever talk to us.
screening · Rahul Menon
46% evidenced · ceiling 93%
- PythonEvidenced
“Python across the data platform since 2017”
- Advanced SQLEvidenced
“SQL — expert. Warehouse design and optimisation.”
- SnowflakeNot addressed
The CV does not address this — no evidence either way.
- AirflowNot addressed
The CV does not address this — no evidence either way.
- AWSEvidenced
“AWS: S3, Lambda, Glue, EMR”
Our own product
We do not only build software for other people
Claravo is ours. An AI presentation and report maker that checks its own numbers. It writes from your own files, then looks every number up in the source that line cites — and tells you when it is not there.
Built, shipped and run in-house. It is the same engineering you would be buying, on a product you can open in a browser right now instead of taking on trust.
- Found there
- Found elsewhere
- Not found
Every figure gets one of three answers. Nothing gets a score.
How we work
Paid discovery, a plan you keep, then a fixed scope
Four stages, with the terms stated rather than implied. Nothing after the discovery is committed until you have read the plan it produces — and the plan is yours either way, including if you take it to somebody else.
01
Paid discovery
We watch how the process actually runs — not how the process document says it runs — with the people who do it. Usually one to two weeks. The fee is fixed and agreed before it starts.
You end up with: A map of the process as it really works, an architecture, and a costed plan for building it.
02
A costed plan
The plan is priced from what discovery found rather than estimated from a conversation, so the number you are asked to approve was derived rather than guessed.
You end up with: The plan itself, yours to keep whether or not you continue with us — including if you take it to somebody else.
03
A fixed scope
Work ships in short increments with something you can look at at the end of each one, rather than disappearing for a quarter and returning with a reveal.
You end up with: Working software in your own repositories and accounts, built in increments you can stop at.
04
Handover
Documentation, your accounts, your pipelines, and a build another team could pick up. This is part of the work rather than a favour at the end.
You end up with: Everything needed to run and change it without us. There is no version of this where what we built stops working because you stopped paying us.
Limits
What it will not do without asking you
Every workflow we build sits at one of three levels, chosen deliberately rather than by default — and it only moves up when the measured numbers earn it.
Level 1
AI Suggests
The person decides
AI produces a recommendation, a draft or a shortlist. Nothing happens until someone chooses to act on it.
Judgement calls, anything customer-facing, and every process on day one.
Level 2
AI Assists
The person reviews
AI does the work and presents it complete. A person checks and approves, which takes seconds rather than the minutes doing it would have.
High-volume work where the output is easy to verify at a glance.
Level 3
AI Executes
The rules decide
Actions run automatically when confidence and business rules are both satisfied. Everything else falls back to a person, and every action is logged.
Well-understood, high-frequency steps with a measured error rate.
And where we stop
You are being asked to let a vendor into your systems. That deserves a straight answer in writing rather than a reassuring adjective.
If you are a staffing firm, one more: we never source, submit or place candidates, and we will never compete with you for a placement or a client. Why that limit is the point.
The technical controls, the architecture and what happens when the AI is wrong are on what we build.
Your data stays in your systems
We build inside your infrastructure and your accounts wherever your environment allows it, and we take no copies away.
Nothing crosses between clients
Your workflows and your process improvements are yours. They do not become the next engagement’s head start.
We do not train models on your data
And we do not use your business as a case study without written permission.
We sign your NDA before the detailed conversation
Not after it, and not once the scope is already on paper.
Start small
Find one process worth automating
You do not need to replace your technology stack. Start with one process that costs your team hours every week, and find out what it would take to fix it.

