ONE CUSTOMER · SIX MISSIONS · NINE ARTIFACTS

Forward Deployed Engineer Capstone Project.

SHIP THE MESSY MIDDLE

Northstar Mutual does not need another AI demo. It needs a safe claims workflow that adjusters adopt, leaders can measure and operators can own after you leave.

Start the brief →

0/9artifacts evidenced

01Build, don’t narrate

Every claim needs a repository, deployment, measurement or customer-facing artifact.

02Reveal constraints in order

Make a decision first. Then reveal the customer change and revise your evidence.

03Defend with a person

Self-review prepares you; it does not certify you. Ask a peer or mentor to challenge the trade-offs.

THE ENGAGEMENT

Six staged missions

Suggested pace: 6–8 weeks · 8–12 hours/week

Discover the real workflow

Northstar Mutual receives 1,200 property claims weekly. Adjusters manually triage emailed PDFs and photos; leaders ask for “an AI claims agent.” Find the narrower operational problem.

Discovery notes + stakeholder map

Turn conflicting interviews into a shared picture of users, incentives and risk.

At least three stakeholder viewpoints
Current workflow and failure costs
Unknowns separated from facts

Problem, metric + scope

Define a valuable first deployment and explicitly refuse attractive distractions.

One measurable outcome with baseline
In/out/non-goals
Named owner and decision date

PRACTICE GATEScoping simulator→

Design the safe thin slice

Propose an assistive intake service that extracts claim facts, flags missing evidence and routes uncertain cases to an adjuster.

Architecture decision + threat model

Make the integration, data boundaries and failure choices reviewable before building.

Alternatives and trade-offs
PII/data-flow diagram
Abuse cases and mitigations

PRACTICE GATESecurity review gauntlet→

Integrate and measure

Build the smallest deployed workflow with deterministic validation, human review and observable failure paths.

Working deployed integration

Prove the thin slice works in an environment another person can exercise.

Reproducible setup and seeded demo
Auth, validation and failure handling
Logs, metrics and rollback path

Golden dataset + eval report + CI gate

Replace an impressive demo with repeatable evidence of quality and safety.

Versioned representative cases
Metrics tied to business harm
CI threshold and failure analysis

PRACTICE GATEEval Builder lab→

Survive production reality

Run a failure exercise and communicate impact while preserving evidence for diagnosis.

Incident drill + postmortem

Demonstrate calm diagnosis, customer communication and systemic learning.

Timeline and customer impact
Evidence-backed contributing factors
Owned corrective actions

PRACTICE GATEIncidents, on-call and SLAs→

Prove adoption and value

Compare the deployment against the agreed baseline and recommend what should happen next.

ROI review + adoption evidence

Show whether the deployment changed work—not merely whether the software ran.

Before/after metric with caveats
Usage or workflow adoption evidence
Continue/change/stop recommendation

PRACTICE GATEProving value and ROI→

Transfer ownership

Defend the work, train operators and identify what is reusable for a second insurer.

Handoff runbook + demo video

Let the customer operate, diagnose and safely escalate without the builder present.

Operator-tested runbook
Alerts, ownership and escalation
Demo covers happy and failure paths

Generalisation memo

Separate reusable product insight from one customer’s local workaround.

Reusable vs customer-specific decisions
Next-customer validation plan
Productisation risks

PRACTICE GATEHandoff and enablement→

PROJECT DEFENSE

Score the evidence, then defend it.

0 = absent · 1 = fragile · 2 = portfolio-ready · 3 = exceptional

Problem judgment

Does the work solve a verified workflow problem with explicit trade-offs?

Engineering quality

Can another engineer reproduce, observe, test and safely roll back the system?

AI rigor

Are model quality and safety measured on representative, versioned cases?

Customer delivery

Are adoption, value, communication and ownership transfer demonstrated?

SELF-REVIEW 0/12

Evidence is not defense-ready yet.

Require all nine artifacts and a score of at least 2 in every dimension. In a 15-minute recording: explain the customer problem, demo one happy and one failure path, defend two trade-offs, show the eval and ROI evidence, then describe what you would change.

This is a self-review readiness signal, not a credential or proof of professional experience. Have a real reviewer inspect the linked evidence.