Amara Vale
Madsoft venture · amara-vale.com
A virtual-persona brand run end to end on our own infrastructure: a governed generation pipeline, a brand site, and a companion chat whose safety is decided before the model is called.
One archived identity reference anchors everything generated from it. Candidates are reviewed on a contact sheet, and approving a shot files it under a naming convention, writes the prompt beside it and appends a provenance entry — so every published asset can be traced to what produced it. Image and video run on permissively licensed models on a rented GPU started and stopped by one command, which removes any per-frame dependency on a vendor API. The brand site carries the gallery, a product waitlist and partner enquiries; the companion chat runs on Claude behind an adult gate, an AI-disclosure acknowledgement, deterministic minor detection ahead of the model, crisis-phrase handling, and memory the person can read, edit or switch off. Site, chat, mail, DNS and certificates run on one server we administer.
Outcomes
- Every published asset carries its prompt and a provenance entry
- Disclosure, age and crisis rules enforced before the model is called
- Generation, site, chat and mail on infrastructure we run ourselves
Engagement
Client named under NDA. Several engagements are active public-sector contracts, so we describe the mechanism rather than the customer.
Source, references and a walkthrough can be arranged for serious evaluation.
Synthetic media is trivial to produce and almost impossible to govern. The hard part was never generating an image; it is being able to say, a year later, exactly what produced the one you published — and running a product that holds conversations with people without a vendor's content policy as the only thing standing between them and harm.
A persona has to stay recognisably one person across hundreds of images and months of posting, which no amount of prompting achieves on its own. It is consumer-facing, so disclosure and age rules are a floor rather than a feature. And the economics do not survive paying an API for every frame generated, most of which are rejected.
- 01
One identity reference, then approvals
Consistency is a filing problem before it is a model problem. A single archived master reference, never published, anchors every generation, and approved shots become the anchors for the next ones. The pipeline reads its prompts from the same documents a person edits, so changing the shot list changes what is generated — there is no second copy to keep in step.
- 02
Provenance written by the tool, not by discipline
Approving an asset is one command, and that command does the filing: the conventioned name, the prompt stored beside the file as a sidecar, and a line appended to the provenance log. A record that depends on somebody remembering is a record you do not have on the day you are asked for it.
- 03
Safety before the model, not inside the prompt
Age and crisis handling are deterministic code that runs before a request ever reaches the model: a stated minor blocks the account outright, crisis phrases surface help and change what the model is told, and limits are enforced per account. A system prompt is guidance. A gate is a gate. Nothing is monetised through dependency, and that is a product rule, not a setting.
- 04
Own the pipeline and the metal
Image and video models are permissively licensed and run on a GPU we rent by the hour, and the site, chat, webmail and certificates sit on a single server under our administration. A brand whose face is generated by somebody else's API, and whose audience lives in somebody else's inbox, has outsourced the only asset it has.
The technical dossier for this engagement
This page describes the shape of the work. The dossier describes how it was actually built — and goes to a named person at a named institution rather than onto a public URL.
- System architecture, including the integration boundaries and why they were drawn there
- The data model for the records the system is accountable for
- Security posture: authentication, authorisation, audit trail and key handling
- Deployment topology, failover behaviour and what happens during an outage
- The measured before-and-after, where the client has cleared it for release
- Named CVs for the engineers who did the work
Sent under a mutual NDA. Client names and locations are redacted unless that client has given written consent, which some have and some have not. Prepared individually — expect it within two working days, not instantly.
Next project
Institutional Records Platform
Start here
Bring us the problem nobody has solved.
A 30-minute technical briefing. We will tell you what we would build, roughly what it costs, and honestly whether we are the right people for it.
- ask@madsoftech.com
- Telephone
- +220 872841578
- Office
- Brusubi, The Gambia