We don’t skip stages.

Not once.

§

Northwind runs every engagement through the same five-stage process, from readiness assessment to production operation. It is slower than a hackathon build, and it is why our systems are still running six months later.

Assess

2–5 weeks

Design

1–2 weeks

Build

4–12 weeks

Deploy

1–2 weeks

Operate

Ongoing

Trusted in production by

Meridian Trust

Alderleaf

Cobalt Mutual

Ferrovia

Northfield & Cole

Hearn Whitby

01 · The process, in full

Five stages.

Every engagement, no exceptions.

This is what actually happens between “we signed the contract” and “it’s running in production.”

i

2–5 weeks · Fixed fee

Assess

Readiness, data audit and use-case prioritisation. Ends in a roadmap you can fund, and sometimes in a clear “not yet”.

Use cases scored on one matrix

Typ. 15–25

Roadmap with budget bands

12 months

ii

1–2 weeks

Design

Architecture, evaluation plan and guardrails defined before a line of production code. The go / no-go metric is signed here.

Gold evaluation set

200+ cases

Human review thresholds

Per field

iii

4–12 weeks

Build

The system, integrated into the tools your team already uses, scored against the gold set every week so nobody is surprised at launch.

Weekly eval runs

Every Friday

Code in your repositories

From day one

iv

1–2 weeks

Deploy

Shadow mode against live traffic, then production, monitored from the first request rather than bolted on after an incident.

Shadow mode

2–3 weeks

One-click rollback

< 5 min

v

Ongoing

Operate

Drift monitoring, scheduled retraining and a clear handoff to your team’s ownership, or a Retainer if you would rather we keep running it.

Drift monitoring

Continuous

Quarterly review

With the builders

The platform

Six layers.
One platform.

Everything an AI product needs, built to work together.

  1. Connect every source in minutes, with lineage on every field.

03 · What each stage draws on

No vendor lock-in

The same eight capabilities,

recombined for your workflow.

01

Language model orchestration

Routing, retries and cost ceilings across models.

02

Retrieval & vector search

Grounded answers with citations you can audit.

03

Agent frameworks & tool use

Typed tools, human checkpoints, full traces.

04

Document extraction & OCR

Field-level confidence, not page-level guesses.

05

Forecasting & time series

Models that report their own drift.

06

MLOps

Training, deployment and monitoring as code.

07

Data pipelines & warehousing

Versioned sources with lineage.

08

Evaluation & guardrails

Gold sets, red-teaming and release gates.

05 · The process, applied

Same five stages.

Four different industries.

64

Systems in production

$41M

Value, trailing 12 months

11 wks

Median to launch

0

Stages skipped

01

Meridian Trust Bank

Financial Services

Avg. loan-file review

1.5 h

Was

4.1 h

−63%

Loan-file review cut 63% with agents

that know when to stop.

Underwriters were re-keying the same fields across four systems. An agent now does the cross-checking and flags only what needs a human decision.

$2.1M

Annual cost avoided

0.62

Review threshold, conf.

9 wks

Kickoff to production

Read the case

01

Meridian Trust Bank

Avg. loan-file review

1.5 h

Was

4.1 h

−63%

Loan-file review cut 63% with agents

that know when to stop.

Underwriters were re-keying the same fields across four systems. An agent now does the cross-checking and flags only what needs a human decision.

$2.1M

Annual cost avoided

0.62

Review threshold, conf.

9 wks

Kickoff to production

Read the case

02

Alderleaf Health Network

Healthcare & Life Sciences

Straight-through intake

91%

Was

0%

+91 pts

Clinical intake that clears itself

before staff sit down.

Referral letters, lab results and insurance forms are extracted, validated against the patient record and routed across 14 clinics.

4,200 h

Staff hours saved / month

14

Clinics onboarded

14 wks

Kickoff to production

Read the case

02

Alderleaf Health Network

Straight-through intake

91%

Was

0%

+91 pts

Clinical intake that clears itself

before staff sit down.

Referral letters, lab results and insurance forms are extracted, validated against the patient record and routed across 14 clinics.

4,200 h

Staff hours saved / month

14

Clinics onboarded

14 wks

Kickoff to production

Read the case

03

Cobalt Mutual Insurance

Insurance

Avg. claims handling

13 min

Was

21 min

−38%

A claims assistant that answers

with its sources.

Adjusters ask in plain language and get answers grounded in policy wording and past decisions, each one cited.

+27 pts

First-contact resolution

94%

Grounding accuracy

10 wks

Kickoff to production

Read the case

03

Cobalt Mutual Insurance

Avg. claims handling

13 min

Was

21 min

−38%

A claims assistant that answers

with its sources.

Adjusters ask in plain language and get answers grounded in policy wording and past decisions, each one cited.

+27 pts

First-contact resolution

94%

Grounding accuracy

10 wks

Kickoff to production

Read the case

04

Ferrovia Logistics

Manufacturing & Logistics

Forecast error (MAPE)

12.5%

Was

18.9%

−6.4 pts

Forecasts that say when

they are losing accuracy.

Demand models across 22 distribution centres, monitored for drift and retrained on a schedule the planners can see.

−22%

Safety-stock cost

22

Warehouses

16 wks

Kickoff to production

Read the case

04

Ferrovia Logistics

Forecast error (MAPE)

12.5%

Was

18.9%

−6.4 pts

Forecasts that say when

they are losing accuracy.

Demand models across 22 distribution centres, monitored for drift and retrained on a schedule the planners can see.

−22%

Safety-stock cost

22

Warehouses

16 wks

Kickoff to production

Read the case

06 · Where to enter the process

Fixed fees

Enter at any stage.

Most start at the first.

Assessment

Know what to build

Teams with an AI budget and no shared way to prioritise between use cases.

From

$18k

2–5 weeks

Data & infrastructure audit

Use-case matrix: effort, risk, payback

Build-vs-buy per use case

12-month roadmap with budget bands

Pilot

Most popular

Prove it works

Teams with a specific use case who need to prove it before a full build.

From

$65k

6–10 weeks

One system, scoped to ship fast

Evaluation harness before launch

Guardrails and human checkpoints

A clear go / no-go on scaling

Scale

Run it for real

Teams taking a proven use case to full production volume.

From

$180k

3–6 months

Full engineering team

Monitoring and alerting from day one

Integration with systems of record

Runbooks and an agreed handoff

Retainer

Keep it accurate

Teams who want us to keep operating a system after launch.

From

$12k

per month

Drift detection and monitoring

Scheduled eval and retraining

Incident SLA

Quarterly roadmap review

07 · In their words

From people who sat through every stage

Renee Castellano

Meridian Trust Bank

Dr. Aaron Mbeki

Alderleaf Health Network

Lindsay Oduya

Cobalt Mutual Insurance

Tom Baptiste

Ferrovia Logistics

Whitney Sarr

Orinda Freight Co.

“

We didn’t want a black box making credit decisions, and Northwind didn’t build us one. The agent does the tedious cross-checking and tells us exactly why it flagged what it flagged. Our underwriters trust it because they can see its work.

Renee Castellano

VP of Loan Operations · Meridian Trust Bank

−63% review time

08 · Start

Start with

stage one.

A 30-minute readiness call is the first hour of the Assess stage. You keep what comes out of it either way.

Create a free website with Framer, the website builder loved by startups, designers and agencies.