Talk to us
Selected work Anatomy of an agent Capabilities Principles Contact
tech@consolider.africa +27 81 049 5520 Cape Town ·

Neuro · Applied AI · Cape Town

Intelligence, engineered.

AI agents, automation, models and data, built to do real work inside real operations. Not demos.

Start an AI project How an agent works

01 why neuro

Most AI never leaves the demo. We build the kind that clocks in: agents wired into your systems, grounded in your own data, tested against real work and looked after once they are live.

02 anatomy of an agent

Five steps between a request and a result.

  1. 01 · input

    Listen

    A message, a document, a form, a ticket or a schedule. The agent starts where the work already arrives.

  2. 02 · retrieve

    Remember

    It pulls the right context from your files, databases and systems, so answers come from your facts, not guesses.

  3. 03 · reason

    Decide

    A model plans the next move inside rules you set: what it may do alone, and what needs a person.

  4. 04 · act

    Do

    It uses real tools: updates the CRM, drafts the reply, files the record, books the slot, triggers the workflow.

  5. 05 · verify

    Check

    Every action is logged and checked. Low confidence goes to a human, and the system learns from the correction.

03 selected work

Agents already at work.

Five AI systems we built, taken from our own code. Public products are shown live. Private systems are shown from local builds with blank data, so no customer record ever appears on screen.

Fourcee website on a MacBook: the missed-call section
9:41
Fourcee on iPhone: the 3D ring builder
9:41
Fourcee on iPhone: the call-back feature with a live outreach card
Live site, captured at native size
  • MacBook fourcee.online · missed-call audit
  • iPhone 3D ring builder
  • iPhone call-backs
01Live

Fourcee

Voice AI receptionist for jewellers

Independent jewellers miss calls while they serve the counter, and a missed call can be a lost sale or repair.

Live
What we built
A phone agent for each store, created through the ElevenLabs agent API on a Twilio number. It says up front that it is an AI, answers from a knowledge base built from the store’s hours, services and FAQs, takes enquiries, requests bookings and callbacks, and transfers to staff. It flags and ends calls that probe for security details. The public site adds a free missed-call audit and a 3D ring builder.
Stack
  • ElevenLabs agents
  • Twilio
  • Node
  • React 19
  • Three.js
fourcee.online
Neuro Ops agent hub on a MacBook, local build with blank data
9:41
Neuro Ops hub on iPhone, local build
9:41
Neuro Ops insights on iPhone, local build
Local build · blank data · no customer records
  • MacBook agent hub + voice composer
  • iPhone hub
  • iPhone venture insights
02Running in-house · private

Neuro Ops

Our own AI operations control plane

Running several ventures from one desk means the same checks, chasing and reporting every week.

What we built
A hub of seven AI agents, each with its own role and voice, plus a Voice Room. A router decides what a spoken request needs before any chat model sees it: an app idea goes to the App Factory, an image or 3D job goes to a worker, a status question goes to the job ledger. Anything that would merge, deploy, send or spend becomes a card for a person to approve. Voice can never approve.
Stack
  • Node
  • Express
  • PostgreSQL
  • Docker
  • ElevenLabs
  • Vercel AI Gateway
Private system · screens from a blank-data build
App Factory on a MacBook, empty state before the first scout run
9:41
App Factory on iPhone, empty idea queue
9:41
App Factory on iPhone with the navigation drawer open
Local build · blank data · no customer records
  • MacBook idea queue · empty, awaiting first scout
  • iPhone queue
  • iPhone navigation
03In build · private

App Factory

AI idea scout with a research gate

App ideas are cheap. Knowing which ones deserve to be built is the hard part.

What we built
An AI scout proposes app ideas. Deterministic rules reject duplicates and generic patterns. A research step that respects robots.txt collects demand and competitor signals from Hacker News, the App Store, GitHub, Product Hunt and Google Trends, then gives a weighted score out of 100. To pass, an idea needs at least 70 overall, minimums on every dimension and five signals from two or more sources. A person reviews everything that passes, and publishing is signed and behind a kill switch.
Stack
  • AI agents
  • Node
  • Playwright
  • Weighted scoring
Private system · screens from a blank-data build
Agent Farm on a MacBook, waiting for the first machine snapshot
9:41
Agent Farm on iPhone, live screens waiting for first still
9:41
Agent Farm machines list on iPhone, no data yet
Local build · blank data · no customer records
  • MacBook farm overview · no snapshot yet
  • iPhone live screens
  • iPhone machines
04In build · private

Agent Farm

Coding agents across four machines

Coding agents are fast. They should not ship anything without a person signing off.

What we built
An orchestrator spread across four PCs: the control plane on one, and engineering, marketing and VR lanes on the others. A model rates each request for difficulty, lane, urgency and risk and turns it into a task card with acceptance checks. Nothing runs without an owner OK. Runners can only open draft pull requests. A shared skills library promotes a skill from unproven to proven only on evidence.
Stack
  • Vercel AI SDK
  • Node
  • GitHub
  • Syncthing
Private system · screens from a blank-data build
9:41
9:41
Illustrated · the prototype runs on sample data only
  • Illustration capture → review
  • Note no sample data shown
05Prototype · private

Receipt reader

Vision model for crew expenses

Crew expenses arrive as photos of paper receipts, and someone has to type them in.

What we built
A Supabase edge function sends a receipt photo to a vision model and gets back the supplier, date, amount, currency and category, with a confidence score. Low-confidence reads are flagged for review, and nothing is saved until a person confirms. It is part of a yacht operations prototype that runs on sample data only, so there is no public showcase.
Stack
  • OpenAI vision
  • Supabase Edge Functions
  • TypeScript
Private prototype · illustrated, not screenshotted

04 capabilities

Six things we build.

Assistants that read, decide and act across email, documents, CRMs and internal tools, with guardrails and a person approving what matters.

  • Agent design
  • Tool use
  • Approval queues
  • Audit trails

Phone and in-browser voice that answers, books and hands over cleanly, built on telephony and natural speech rather than menus.

  • Inbound calls
  • Booking
  • Call handover
  • Speech pipelines

The repetitive steps between your systems, removed: intake, routing, data entry, reporting and follow-ups.

  • Intake & routing
  • Scheduled runs
  • Reporting
  • Integrations

Search and question-answering over your own knowledge, with sources shown, so teams can trust what comes back.

  • Knowledge search
  • Source citations
  • Document pipelines
  • Access control

Choosing, prompting, fine-tuning and testing models against your real tasks before anything reaches production.

  • Model selection
  • Prompt design
  • Eval sets
  • Cost tuning

Products and internal platforms with AI built in from day one: copilots, dashboards that explain themselves, plain-language interfaces.

  • Copilots
  • Dashboards
  • Plain-language UI
  • Product builds

05 principles

Four rules we ship by.

01

Grounded

Answers come from your data, with sources you can check.

02

Measured

We test against real examples of your work before launch, and keep testing after.

03

Supervised

People stay in control. You decide what the system may do alone.

04

Yours

You own the code, the prompts and the data. No lock-in to one model provider.

06 process

From idea to in production.

  1. discover

    Find the work worth automating and define what success looks like.

  2. prototype

    A working version on real data, fast, so you can see it think.

  3. evaluate

    Measure it against your examples. Fix what fails. Repeat.

  4. deploy

    Wire it into your systems with logging, permissions and fallbacks.

  5. operate

    Monitor, improve and adapt as your business and the models change.

07 the group

Part of Consolider.

Parent · Software solutions CONSOLIDER

Custom software, AI and VR integration, 3D web. Trading as 9 Lives.

Visit Consolider ↗
Applied AI

Models, agents, automation and data. Intelligence, engineered for real operations.

You are here
Mechatronics

Robotics, embedded systems, control and software. Machines that move with purpose.

Visit Paraboliq ↗

08 contact

Give it a job.

Tell us the task that eats your team's week. We'll tell you honestly whether AI should do it, and how.

Email
tech@consolider.africa Phone
+27 81 049 5520 WhatsApp
Message us