Available for AI automation builds

Northscale Studio

AI automation studio

"If the work repeats, the system should run it. If AI acts on it, humans should be able to govern it."

We build AI automation systems for businesses: prompt-engineered agents, workflow dashboards, email and call automation, qualification flows, RAG, security, governance, and audit trails.

View our work Start a project
Live Gated RAG
AI Agents
Prompt Engineering
RAG
Security
Governance
Workflows
Supabase
Trigger.dev
Next.js
Claude
Gemini
OpenAI
Approvals
AI automation first

We build AI workflows that take real work off your team.

Northscale Studio designs AI automation systems for businesses that need more than a chatbot. We build the prompts, agents, workflow logic, dashboards, integrations, and rules that make AI useful in daily operations.

A system can qualify leads, draft emails, route support requests, prepare call notes, monitor social workflows, and hand risky decisions back to a human. RAG, security, governance, approvals, and audit trails are built in.

Multi-agent Specialized agents for calling, email, social, qualification, reporting, research, and routing.
Prompt-engineered Clear system prompts, tool rules, scoring rubrics, escalation paths, and output quality checks.
RAG-ready Business knowledge connected to AI through memory, retrieval, and structured context.
Governed Approval gates, audit trails, permission checks, and human review for risky actions.
Observable Heartbeats, statuses, incidents, costs, and run history visible from one place.
Specialists

Small team. Senior automation focus.

Northscale Studio is led by specialists who design AI systems around real business operations, from agent workflows and prompt engineering to governed dashboards and implementation.

AI agent engineer · voice & operations

Alwyn van Vuuren

Operations-focused systems builder with years of experience in freight forwarding and logistics. Focused on identifying operational bottlenecks and designing practical AI-assisted workflows that reduce manual work and improve efficiency.

Software developer · AI operations

Marelize Claassens

Software developer focused on AI operations, workflow automation, and scalable business systems. Background in Information Systems with hands-on experience building governed AI workflows, operational tooling, and multi-agent platforms.

What we build

Services

If a process repeats, we can build an agent for it. These are the pillars — the agent itself is always built around your operation.

01

Custom AI agents

Any agent your operation needs — voice, email, chat, research, qualification, scheduling, reporting, data entry. Built to your workflow, never forced into a template.

Get a quote 02

RAG knowledge systems

Your documents, SOPs and decisions turned into trusted retrieval, so agents act on real context instead of guessing.

Get a quote 03

Agent platforms & governance

Operations dashboards, multi-agent orchestration, approval gates, audit trails and cost control — the rails that make AI safe to run.

Get a quote
Selected work

Recent work

Live production systems — voice agents, repair triage, email & recruitment screening, website qualification and logistics planning — plus Cortex, the operational platform we're building to run them all.

Independent build · live
"Call in, describe the fault, get the fix."

Repair agent · appliance voice triage

TwilioOpenAI RealtimeClaudeSupabasepgvectorRAGFirecrawl

The problem

Appliance brands burn truck-rolls and customer time on warranty calls the user could self-resolve with reliable step-by-step guidance — and a safety screen before any guidance is given.

What we built

A callable phone line that runs a safety pre-screen, triages the fault by voice, retrieves over 304 verified manufacturer repair guides (889 embedded chunks) via hybrid pgvector + tsvector RRF, walks the caller through the fix, and books an engineer when DIY is unsafe or out of scope.

What it proves

A live insights dashboard, dispatch SMS straight from the call, and a full transcript and audit trail per call. EU AI Act disclosure and the recording notice are baked into the opener.

71%self-resolution rate
304verified repair guides
889embedded chunks

Live · phone demo plays the agent's actual voice · dashboard from production

In active development
"The operational layer for AI-powered businesses."

Cortex · operational intelligence platform

Multi-agent coordinationApprovals & governanceAudit & observabilityOperational memoryPolicy & risk tiersEach client isolated

The problem

Every business is about to run AI workers — and almost none are prepared to manage them. Most agents answer, act, then forget. Across a real operation you need memory, governance, approvals, escalation and audit, and a way for the whole system to get smarter over time, not just faster.

The operational layer

Cortex runs, governs and improves every AI worker in one place: approvals, observability, traces, evaluations, policy enforcement, risk tiers, scoped tool permissions and a credential vault — each client isolated.

Operational memory

The leap: Cortex remembers, observes and evaluates across time, then improves the workflows behind your AI workers — so the longer they run, the more reliable and accurate they get. Most platforms just run agents; Cortex learns.

1control plane · every client isolated
Improvesgets better the longer it runs
20+industries with tailored templates

In active development · governed-by-design, each client isolated · a live preview you can click is coming soon

Production · live
"From ring to resolved without a human picking up."

Service business voice agent

VapiTwilioClaude Sonnet 4.6ElevenLabsDeepgramSupabaseTrigger.dev

The problem

Salons, dentists, plumbers and other service desks lose hours to bookings, reschedules and repetitive FAQs, with no structured record of what was actually said or promised on the call.

What we built

A Vapi-orchestrated voice agent: Twilio inbound, Claude Sonnet 4.6 reasoning over a typed tool schema, ElevenLabs streaming TTS at ~250 ms TTFB, Deepgram STT, every booking written atomically into Supabase. Post-call summarisation runs in Trigger.dev.

What it proves

Sub-second turn-taking, deterministic tool calls, atomic upserts on the call ID, an audit trail per call, and human escalation when intent slips. In production at a hair salon on the real stack.

~250 msTTS time-to-first-byte
Sub-1sturn-taking latency
100%calls with audit trail

Live · plays the agent's actual production voice · real transcript

Live · in-tenant
"Inbox zero, every day. Without lifting a finger."

Outlook · email triage agent

Microsoft GraphOutlookExcel RESTSharePointClaude Haiku 4.5Claude Sonnet 4.6Vision extraction

The problem

Knowledge workers spend 28% of every week reading, sorting and forwarding email. Most of it is process, not thinking. Reply lag costs deals, buried attachments stall reporting, and misfiled threads break the audit trail.

What we built

An autonomous triage layer inside Microsoft 365. It reads every inbound mail, classifies intent, files it, drafts the reply in the user's tone, extracts PDFs and updates the SharePoint Excel workbook that owns the data, books meetings end-to-end, and surfaces only the threads that need a human to decide.

What it proves

Runs entirely in-tenant with EU residency, a full audit ledger, and a per-user kill switch.

94%auto-handled rate
11.4hreclaimed / user / week
18smedian draft latency
99.3%PDF extraction accuracy

Live · in-tenant Microsoft 365 · PDF → Excel extraction proven in production

Production · adopted internally
"Every container, predictable to the hour."

Freight inbound planning dashboard

ViteVercel FunctionsSupabase · PostgresLive AISInland gaugesFirecrawl

The problem

Inland container planning is a coordination problem. Vessel cutoffs, customs windows, waterway levels, truck capacity, barge slots, terminal rules and rolled-container risk all live in different tools. Every tab switch is a chance to roll a container and break a promised delivery date.

What we built

Eight planning surfaces on one operating canvas — export, import, CY/CY, terminal directory, live port activity, network schedules, waterway notices, vessel admin — against 32 inland terminals and 12 live data feeds.

What it proves

Significant on-time-delivery improvement, material reduction in container downfall, and planning time per booking cut roughly 5×. Routing decisions log modal split and CO₂ score natively for CSRD scope-3 reporting.

~5×faster planning per booking
32inland terminals covered
12live data feeds

Built independently · adopted internally by a freight inbound operations team · real port AIS · real inland waterway gauges

Independent build
"Click the orb. Talk. Book."

Website voice agent

VapiClaudeWeb SDKReal-timeTwilio SMS

The problem

Static contact forms make prospects do the work. By the time a form is read, the buying moment is often gone.

What we built

A tap-to-talk website agent that introduces the service, answers questions live, qualifies the visitor, and books a discovery call — the same agent you can talk to on this page.

What it proves

Turns inbound intent into a structured conversation, then routes the transcript, summary and contact details to the right place instead of a dead inbox.

Independent build · live on this site · transcript + summary routed on every call

Production notes

What we believe about automation

Automation is not a prompt box

The value is the system around the model: data, permissions, workflows, review, memory and observability.

Governance is a product feature

Approval gates, audit trails and scoped tools are what make teams trust AI with real operational work.

Humans govern judgment, agents handle repetition

Agents run the repeatable work. People own the decisions, approvals and final accountability.