Skip to content
revot
revot

Software & automation studio

Software, shaped around how you work.

Revot designs and builds practical software and automation around the way your business really runs, from the systems your customers touch to the workflows behind them.

Companies we've worked with.

  • Vocallio
  • DigitalAtlas
  • Saphen
  • Darul Huda
  • Golden Dragon
  • Money Saby
  • Synthicare
  • Yafi.ai
  • Votours
  • New Hope Wellness
  • FMFlow
  • FirstandKing
  • Anzeri AI
  • Ezbikes
  • Convoly.ai

The problem

Most operational cost hides in the handoffs.

Rarely is one big thing broken. More often it is a dozen small handoffs that each depend on someone remembering, copying or chasing. Each one is cheap on its own. Together they decide how fast you answer customers and how much your team can carry.
  1. 01

    The same details, typed three times

    An enquiry arrives by email, gets copied into a spreadsheet, then re-keyed into the CRM. Every copy is a chance for a typo and a reason for a delay.

    re-keying

  2. 02

    Calls nobody is free to answer

    Phones ring during busy hours and after closing. Routine questions tie up skilled people, and what was said on the call rarely reaches the system that needs it.

    missed calls

  3. 03

    Answers that live in a few heads

    Policies, prices and procedures are spread across documents, recordings and old threads. New staff ask the same people the same questions.

    tribal knowledge

  4. 04

    Approvals waiting in an inbox

    A decision needs sign-off, so it waits. Nobody can see where it is, and the customer asks again.

    approval limbo

  5. 05

    Failures nobody notices

    An integration quietly stops syncing. The first alert is a customer complaint, and nobody is sure which records were affected.

    silent failure

  6. 06

    Tools that disagree with each other

    The phone system, the CRM and the spreadsheet each hold a slightly different version of the truth — so people trust none of them.

    no single record

Each of these is a workflow problem before it is a software problem. That is why we map the work before we write the code.

Selected work

Engagement types, described honestly.

Client details are anonymised. We describe the shape of each problem and what was built — not names, figures or outcomes we cannot yet publish.
Status: BuildCustomer operations

Bilingual voice system for a customer-operations team

Calls needed to be handled conversationally in two languages, without building and maintaining a separate call flow for each. For a voice agent to feel usable, the pause between a caller finishing and hearing a reply has to be short — so speed was a design constraint, not a nice-to-have.

The system sits on the phone line in front of the team. Each turn was broken into stages — end-of-speech detection, transcription, reasoning, speech synthesis and playback — so each could be measured and improved separately rather than guessed at as a whole.

Fig. W1Anatomy of one conversational turn, in either language
Anatomy of one conversational turn, in either languageOne conversational turn: the caller speaks, the end of their speech is detected, streaming speech-to-text produces a transcript in the caller's language, a language model reasons over it, speech is synthesised in the same language and played back. Each stage is measured separately because latency is the gap between the caller finishing and hearing a reply.next turn — caller may interruptCaller speakslanguage a or bEnd of speechdetectedTranscriptstreaming sttReasoningllmSpeechsame languagePlaybackcaller hearsA each stage timed on its own: the engineering work went into the gaps between them
Status: BuildTherapy practice

Privacy-conscious notes application for therapy practitioners

Session notes are among the most sensitive records a practitioner keeps. General-purpose note tools are not designed around who should — and should not — be able to read them.

The design is organised around a trust boundary between the practitioner and the hosted service, with only the minimum necessary data crossing it. We describe this as privacy-conscious design, not a guarantee: the precise properties depend on the threat model, and we share them in detail on request.

Fig. W5A notes workflow designed around a trust boundary
A notes workflow designed around a trust boundaryInside the practitioner's boundary, the practitioner writes and reviews a session note in a note editor. Only the minimum necessary data crosses to the hosted service, which stores protected records and keeps an access log. The precise security properties depend on the threat model and are shared on request.PRACTITIONER SIDEHOSTED SERVICEminimum data crossesPractitionersigned inNote editorwrite + reviewProtected recordsaccess-controlledAccess logwho · whenA design properties depend on the threat model — shared in detail on request
Status: BuildVoice-agent operations

Operations dashboard for a voice-agent service

Running voice agents for several customers means juggling accounts, usage and call history across a platform designed for one. An operator needed a single white-label view over all of it.

The dashboard sits between the operator and the underlying voice platform’s API. Prompt changes are drafted with assistance but published only after a person has reviewed them.

Fig. W2One operator view over many voice-agent accounts
One operator view over many voice-agent accountsAn operator uses one dashboard to manage sub-accounts, a credits ledger and call logs. Prompt changes are drafted with assisted enhancement, reviewed by a person, and only then published to the underlying voice platform.publishusageOperatorwhite-labelDashboardone viewSub-accountsper customerCreditsusage ledgerCall logssearchablePrompt editorassisted draftsReviewbefore publishVoice platformvia its api

Every engagement on this page is anonymised: no client names, figures or locations. Statuses describe the kind of work, not a verified production outcome. Fuller case studies are shared on request and published as clients approve them.

All anonymised work

Process

Map it. Scope it. Build it in pieces you can check.

Every engagement moves through the same four stages. Each one ends with something you can review before the next begins.
  1. Stage 01

    Discovery

    We map the work from trigger to completion.

    Who is involved, which systems, what data, where the handoffs and exceptions are, how often it happens, what it costs in time, and what privacy or retention constraints apply. Then we agree what success looks like.

    You receive

    A workflow map and success criteria

  2. Stage 02

    Proposal

    A scope you can hold us to.

    The problem, the proposed workflow, deliverables and exclusions, milestones, acceptance criteria, dependencies, who is responsible for what, change control, support and data assumptions — written down before work starts.

    You receive

    A written scope with acceptance criteria

  3. Stage 03

    Delivery

    Built in increments you can review.

    Access controls are set up first. We confirm requirements, build in reviewable increments, demo regularly, and test the edge and failure cases — not just the happy path. Deployment and known limitations are documented as we go.

    You receive

    Working increments and regular demos

  4. Stage 04

    Handover & support

    You leave with the keys and the manual.

    Acceptance against the agreed criteria, documentation, secure transfer of credentials and agreed assets, training, and the support arrangement we agreed. Access we no longer need is revoked; data export or deletion obligations are met.

    You receive

    Documentation, transferred assets, agreed support

Reliability & security

Accuracy is designed in, not hoped for.

Software that touches customers, money or sensitive records has to fail safely. These are the engineering habits we bring to every build — and the checks we run before an AI component goes anywhere near production.

Spec. A — Engineering baseline

  • Separate environments

    Development, staging and production are kept apart, so testing never touches live data.

  • Secrets and least privilege

    Credentials live in secret management. Every account and service gets only the access it needs.

  • Validated boundaries

    Inputs and webhook signatures are checked where data enters the system.

  • Idempotent workflows

    Payments and workflow steps are safe to retry — nothing is sent or charged twice.

  • Traceable failures

    Structured logs and correlation IDs let us follow one request through every system it touched.

  • Defensive integrations

    Timeouts, retries and rate limits on calls to services we do not control.

  • Failure cases tested

    We test what happens when things break — not only when everything works.

  • Versioned prompts and evals

    Prompts, tools and evaluation sets are versioned like code, so changes can be compared and reversed.

  • People over consequential actions

    Anything that moves money, changes records or speaks for you can require human approval.

  • Retention, deletion, export

    How long data is kept, and how it is exported or deleted, is documented — not left implicit.

Spec. B — Before an AI component ships

  1. T.01Correctness

    Is the answer right for a set of real questions with known answers?

  2. T.02Groundedness

    Is every claim supported by a retrieved source?

  3. T.03Source quality

    Are the retrieved passages the right ones, and current?

  4. T.04Refusal & escalation

    Does it decline, or hand over to a person, when it should?

  5. T.05Prompt injection

    Can instructions hidden in documents or messages change its behaviour?

  6. T.06Tenant isolation

    Can one customer’s data ever surface in another’s answer?

  7. T.07Latency

    Is it fast enough for the way it is used — on a call, in a chat, in a batch?

  8. T.08Cost

    What does each answer, call or run cost at realistic volume?

  9. T.09Regression

    Did a model or prompt change make anything worse?

Start a project brief

Tell us about one process.

Eight short questions, one at a time: how the work happens today, what breaks, and what good would look like. So the first conversation starts from your problem, not our sales deck.

  • A few minutes. Your answers are saved as you go.
  • We reply with how we would approach it.
  • No sales call needed to get a first reply.

Step 1 of 9 — Costly process

  1. Step 1: Costly process (current)
  2. Step 2: How it works now
  3. Step 3: How often
  4. Step 4: When it fails
  5. Step 5: Tools involved
  6. Step 6: What success looks like
  7. Step 7: Deadline
  8. Step 8: Your details
  9. Step 9: Review

What process costs you the most time or money?

In your own words. One or two sentences is plenty — the next questions fill in the detail.