04 / PATTERNSThree patterns · three sectors · pre-first-engagement

Reference
architectures.

No fabricated outcomes. Each pattern is the anatomy of how we'd build — ingest to handoff — with the risk controls and KPIs we'd instrument from day one.

01 / PATTERN · eGov
Blueprint · ready for first engagement

Agentic permit intake.

A reference architecture for drafting, citing, and routing permit applications with a guardrailed retrieval agent — officer authority preserved, every decision audit-defensible.

ANATOMYIngest to handoff, in order
  1. 01

    Ingest

    Statute corpus, precedent archive, inter-agency data under identity scope

  2. 02

    Retrieve

    Hybrid vector + BM25 over the corpus per intake packet

  3. 03

    Draft

    Model proposes classification + citations; never decides

  4. 04

    Review

    Officer signs off in under a minute per case

  5. 05

    Audit

    Every step trails into an immutable decision log

  6. 06

    Route

    Case handed off to the right department with provenance attached

STACKVendor-agnostic
  • Vector store (chunked statute + precedent)
  • Hybrid retriever (BM25 + dense)
  • LLM orchestration with tool-use
  • Structured-output validation (JSON schema)
  • Audit-log append store
  • Officer review surface
RISK CONTROLSWhat has to hold
Hallucination guards
Refuse-to-answer on retrieval misses; quote-or-abstain policy on citations.
Appeal path
Every automated draft reversible by the officer with one click.
Data residency
In-region storage by statute; zero third-party boundary crossing for PII.
MEASUREMENT PLANInstrumented from day one
KPI

Officer minutes per case

Baseline measured week 1; target −40% by month 3
KPI

Refusal rate on retrieval miss

>95% of out-of-scope queries refused
KPI

Appeal rate on automated drafts

<5% of drafts appealed by officers
ENGAGEMENT SHAPE

4-week advise · 12-week build · 6-month operator handover

02 / PATTERN · Healthcare
Blueprint · ready for first engagement

Ambient clinical documentation.

Reference architecture for ambient scribes — capturing an encounter, producing a reviewer-signoff note, handing off to the EHR with HIPAA-shaped audit. Benchmarked against JAMA 2025.

ANATOMYIngest to handoff, in order
  1. 01

    Capture

    Encounter audio, scoped to consented clinician and patient session

  2. 02

    Transcribe

    On-device or in-region STT with PHI redaction before any remote call

  3. 03

    Structure

    Model assembles SOAP note against institutional templates

  4. 04

    Review

    Clinician forced-choice signoff on every generated field

  5. 05

    Handoff

    Structured note lands in EHR via FHIR or vendor API

  6. 06

    Audit

    Monthly drift audit; PHI retention review on cadence

STACKVendor-agnostic
  • On-device or VPC-bounded STT
  • PHI redaction boundary layer
  • LLM orchestration with institutional templates
  • FHIR / EHR integration adapter
  • Audit store with retention policy
  • Weekly drift-monitoring job
RISK CONTROLSWhat has to hold
PHI leak
Redaction before any third-party boundary crossing; tested on every release.
Clinician burden
Signoff UI measured in seconds, not assumed; friction is a bug.
Model drift
Signed-off notes sampled monthly and compared to ground-truth baseline.
MEASUREMENT PLANInstrumented from day one
KPI

EHR minutes per encounter

JAMA-benchmark 13.4 min reduction (2025 study)
KPI

Signoff completion rate

100% of generated notes reviewed before EHR push
KPI

Drift delta month-over-month

<2% regression; alert if breached
ENGAGEMENT SHAPE

4-week advise · 16-week pilot · quarterly posture review

03 / PATTERN · Oil & Gas
Blueprint · ready for first engagement

Digital twin + predictive maintenance.

Reference architecture for mixed-vintage fleets — sensor fusion into a digital twin, anomaly explanations operators can read, shift-change handover that survives crew rotation. Chevron benchmark.

ANATOMYIngest to handoff, in order
  1. 01

    Ingest

    Sensor data (vibration, thermal, flow, pressure) per asset class

  2. 02

    Twin

    Digital representation of the asset, versioned against engineering docs

  3. 03

    Detect

    Anomaly detection against per-class models

  4. 04

    Explain

    Operator-readable root cause and what changed since last shift

  5. 05

    Handoff

    Shift-change dashboard consumed by oncoming crew

  6. 06

    Close

    Maintenance ticket with provenance and replay

STACKVendor-agnostic
  • Time-series store (asset-partitioned)
  • Per-class anomaly models
  • Digital-twin state store, engineering-doc versioned
  • Explanation layer (root cause + provenance narrative)
  • Operator-first shift-change dashboard
  • Ticketing integration (maintenance system adapter)
RISK CONTROLSWhat has to hold
False positives
Trust score gated; operators override without friction — trust is earned.
Model decay
Re-training cadence tied to asset-class lifecycle, not calendar.
Safety-critical scope
Twin never commands equipment; it only advises and records.
MEASUREMENT PLANInstrumented from day one
KPI

Unplanned-downtime hours

20% reduction (Chevron benchmark, 2025)
KPI

Operator trust score

Monthly survey, positive delta over 90 days
KPI

Time-to-root-cause from alert

<15 minutes, median
ENGAGEMENT SHAPE

4-week advise · 20-week build · 12-month operator handover

10 / ENGAGEMENT

Start with the
outcome that has
to move.

Four-week advise engagement. Fixed scope. A defensible deliverable. If the metric doesn’t want to move, we say so before you sign anything else.

Start an engagement Read the methodology
inquiries@hqscg.comLLC · EST. MMXXIVHQSCG.COM