Skip to contact
AutomateWise
AI, end to end — an independent engineering practice

Automate everythingthat shouldn’t need a person.And nothing that should.

Legacy systems taught to think. Operations that run themselves. Security that keeps pace. We own the whole line — ingestion, models, deployment, monitoring — and we will tell you, in writing, which parts to leave alone.

everything you were soldwhat you actually needed
70M+
events a day, at peak
End to end
ingestion through monitoring
3
clouds, no religion about which
6 wks
to something in production
Legacy modernisationLarge language modelsRAG that retrieves the right thingAgentic workflowsDocument & workflow automationAnti-abuse and fraud MLThreat modelling for AIData pipelinesKubernetesTerraformAWSAzureGCPObservability & SLOsCost engineeringForecastingSearch & rankingNL2SQLEvaluation harnessesOn-call designLegacy modernisationLarge language modelsRAG that retrieves the right thingAgentic workflowsDocument & workflow automationAnti-abuse and fraud MLThreat modelling for AIData pipelinesKubernetesTerraformAWSAzureGCPObservability & SLOsCost engineeringForecastingSearch & rankingNL2SQLEvaluation harnessesOn-call design

On the name

Anyone can automate a process.

Knowing which processes should not be automated — and which should not exist at all — is the part that saves the money.

Most of the value is in the second word.

The situation

You don’t have an AI problem. You have four of them.

They arrive separately, get quoted separately by different vendors, and get blamed on each other.

01

The system nobody wants to touch

It has run the business for a decade and exactly one person still understands it. Every proposal to modernise has been deferred, because the risk of moving is easy to describe and the cost of standing still is not.

02

The work that shouldn’t need a person

Someone re-keys the same fields every morning. Someone reads every ticket to decide where it goes. Someone reconciles two systems that were supposed to talk. None of it is judgement. All of it is salary.

03

The surface that grows every time you ship

Each integration, each model, each helpful answer is another way in. Security review happens at the end — the most expensive possible moment to discover the thing you cannot ship.

04

The pilot that never became a product

It demoed beautifully in March. It is still demoing beautifully. Nobody can say what it costs at real volume, what it does when it is wrong, or who carries the pager.

Four symptoms, one cause: nobody owns the line from your data to the outcome.

We own the line.

The wise part

The cheapest thing we can build you is the thing we talk you out of.

Every item below has been quoted to somebody by somebody. Every one of them we have argued against — often out of our own scope. This is not modesty. It is the whole method.

A fine-tune

A sharper prompt and three good examples. Nine times in ten a fine-tune is an evaluation problem wearing a GPU bill.

A vector database

The Postgres you already run, with an index you already pay for. Add the vector store the week your numbers ask for it, not the week a conference does.

An agent

A queue and a scheduled job — which has never once hallucinated a refund.

A multi-region architecture

One region, done properly, with a restore you have actually tested. Come back when your customers are in the second one.

Automating it at all

Deleting step three. It is astonishing how often step three was the entire problem.

We have run systems at seventy million events a day.

That is exactly why we know what your ten thousand doesn’t need.

What we do

Five disciplines, one team, one invoice.

This is the whole menu — no partner network, no subcontractors you meet after signing. Everything on it we have already shipped at scale, under names you would recognise. Most engagements use two or three of these; almost none use all five, and we will tell you which.

01

Legacy, taught to think

The system you are afraid of is an asset. It has just been waiting for someone patient.

We do not open with a rewrite. We put an interface around what exists, free the data trapped inside it, and move behaviour out one piece at a time — each piece with a rollback and a reason. The business keeps running while it changes underneath.

Assessment & risk mappingStrangler-fig migrationAPIs around the monolithData extraction & modellingIncremental cutover with rollback
02

Applied intelligence

We do not start with a model. We start with the decision it has to make.

Retrieval that finds the right passage rather than a nearby one. Agents with guardrails and an off switch. Evaluation harnesses that tell you the truth before your customers do. Classical models where classical models win — which is more often than the market currently admits.

LLM & RAG architectureAgentic workflows with guardrailsEvaluation & regression harnessesSearch, ranking & recommendationForecasting & classical MLNL2SQL and query understanding
03

Operations, automated

If a person does it the same way twice, it should stop needing the person.

We map the work as it is actually done — not as the process document claims — then automate the mechanical part and leave the judgement with the human, visibly. Success is measured in hours returned, not features shipped.

Document & workflow automationTicket triage and routingBack-office reconciliationHuman-in-the-loop where judgement is realInternal copilots on your own data
04

Security that keeps pace

Every model you ship is another way in. Assume someone is already knocking.

Prompt injection, exfiltration through a helpful answer, poisoned retrieval, abuse at machine speed. We have built these defences at consumer scale and at carrier scale, and we would rather find the hole in week two than in the press.

Threat modelling for AI systemsAnti-abuse & fraud MLRed-teaming LLM applicationsData governance & PII boundariesDevSecOps & supply-chain hardening
05

Ingestion to monitoring

Somebody has to own the boring half. It is the half that decides whether the rest survives.

Pipelines in, models in the middle, telemetry out. Infrastructure as code, releases that roll back, alerts that fire on customer impact rather than CPU, and a cost model you can put in front of a CFO without flinching. Multi-cloud where it earns its keep; one cloud where it does not.

Data pipelines & ingestionAWS · Azure · GCP architectureKubernetes & infrastructure as codeCI/CD and release engineeringObservability, SLOs & on-call designCloud and model cost engineering

Selected work

The names are under NDA. The numbers aren’t.

Six systems we designed, built and kept alive — inside companies whose logos you see every day. Note the spread: intelligence, modernisation, operations, security, reliability. That range is the point. We will walk you through any of them on a call, under NDA, with the architecture on screen.

012025 →Trillion-dollar consumer technology company

Applied intelligence · Security

Twenty-five million decisions a day, no human in the loop

25M+events / day
Problem

Abuse arrived faster than people could review it, and every rule shipped was already out of date.

What we did

Architected and launched an AI decisioning agent from concept to production in under three months, on infrastructure designed to keep deciding when a region does not.

Outcome

Autonomously processes 25M+ events a day and scales past 70M+ across cloud services.

GenAIML decisioningStreamingDistributed systems
022025 →Trillion-dollar consumer technology company

Applied intelligence · Operations

The assistant that learned an entire security organisation

20%faster onboarding
Problem

A decade of heuristics, rules and tribal knowledge lived in the heads of ten people and in wikis nobody could find.

What we did

Built a retrieval assistant over the organisation’s heuristic logic and rule systems, tuned for precision rather than plausibility, with an evaluation set that could prove it.

Outcome

Cut engineer onboarding time by 20% and shortened incident response by making the answer self-serve.

RAGLLMVector searchEvaluation
032025Global marketing & experience cloud

Applied intelligence

Six weeks to the main stage

6 wksconcept → keynote
Problem

An executive commitment to demo a customer-facing assistant at the company’s flagship conference. The date could not move.

What we did

Architected and shipped it end to end in six weeks — retrieval, query rewriting, out-of-scope detection, guardrails — while keeping it production-safe rather than demo-safe.

Outcome

Shipped on the keynote date, then survived contact with real customers afterwards.

LLM assistantRAGQuery understandingGuardrails
042017–2021Enterprise AI platform, Fortune 50

Legacy modernisation · Platform

One stack where nine used to be

9 → 1stacks consolidated
Problem

Language understanding was spread across fragmented, duplicated components that no single team owned.

What we did

Founded and architected a unified, lightweight framework to replace them, and owned the serving infrastructure behind its public APIs.

Outcome

Became the technical foundation for the platform’s language understanding — and earned the company’s Outstanding Technical Achievement Award.

Platform architectureProtocol BuffersML servingREST APIs
052023–2025Global networking infrastructure company

Legacy modernisation · Infrastructure

The migration nobody noticed

Zerodowntime at cutover
Problem

A mixed Linux and Windows estate had grown by acquisition and habit. Every change was a held breath.

What we did

Rebuilt the estate as code — modules, golden images, automated patching — then moved it to hybrid cloud in waves, each wave with a tested rollback.

Outcome

Cutover completed with no customer-visible downtime. Provisioning that took days now takes minutes.

TerraformHybrid cloudLinux + WindowsAutomation
062022–2024Global networking infrastructure company

Operations · Reliability

The night the pager went quiet

−70%out-of-hours pages
Problem

Alerting fired on symptoms rather than on customers. Engineers were woken for things that did not matter and missed the things that did.

What we did

Replaced threshold alerts with SLO-based alerting, instrumented the paths that carry real impact, and rewrote the rotation and runbooks around them.

Outcome

Roughly 70% fewer out-of-hours pages, with faster detection on the incidents that were real.

SLOsObservabilityIncident responseDevSecOps

Client identities are withheld under confidentiality agreements. Figures describe systems the founders personally designed and shipped in their professional capacity. Company names elsewhere on this page refer to individual employment history only.

0M+
events a day at peak
0.00%
the uptime we design for
0+
years of engineering between us
0
disciplines, one team

The team

Two engineers. Twenty years of other people’s hardest problems.

The range above is not a service catalogue assembled to look complete. It is the list of things the two of us have already had to do, at companies where getting it wrong was news.

ex / currently · Apple · Adobe · IBM

The intelligence half

Machine learning architect. Fourteen years turning research into things that run on Monday.

  • ML architect at Apple, on systems that make tens of millions of decisions a day
  • Previously at a Fortune 50 experience-cloud platform, where he founded its first production LLM assistant
  • Founding engineer on a Watson NLP stack at IBM; winner of its Outstanding Technical Achievement Award
  • US patents in AI/ML · UC Berkeley Sutardja Center · published author
ex / currently · Cisco

The systems half

Infrastructure, DevOps and site reliability. The reason any of it is still running next quarter.

  • Infrastructure, DevOps and SRE at Cisco, running Linux and Windows estates at fleet scale
  • Hybrid cloud architecture, infrastructure as code, and automated release pipelines
  • DevSecOps: hardening the path from commit to production
  • Microsoft Certified: Windows Server Hybrid Administrator Associate

On the anonymity

We have not put our names on this yet. The entity is being registered and we are still employed by the companies above, so the practice stays unnamed until the paperwork clears. The names are under NDA. The numbers are not. On the first call you get both of us, cameras on, real names, profiles open.

Straight answers

The questions you were going to ask anyway.

We would rather answer them here than waste the first fifteen minutes of your call.

Because the second word is the one that earns its keep. Automating a process is the easy half and every vendor can do it. Knowing which processes should not be automated, which should be redesigned first, and which should simply be deleted — that is where the money actually is. We named the practice after the harder half.

Start here

Bring us the problem. Not the technology.

No discovery-call gauntlet. No junior account manager. Write us a paragraph and one of the two people who would actually do the work will answer — usually within one business day.

  • A human reads every message. Never a sequence.
  • If the honest answer is that you don’t need us, that is the answer you get.
  • Nothing you tell us leaves the two of us.

Or just email us

hello@automatewise.ai
We reply from a human