Work.

Staff re-keyed data from client-submitted forms, scans, and email attachments into internal systems — hours of transcription weekly, errors surfacing downstream. The documents carry personally identifiable client information under Canadian privacy law, so “just send it to an AI” was never on the table.

An intake assistant that reads incoming forms and PDFs, extracts and validates the fields that matter, flags anything ambiguous for a human, and delivers structured records into the firm's systems. Staff review instead of transcribe.

Processing runs in a Canadian cloud region — data residency preserved. Personal identifiers are detected and masked inside that environment before any external model call; external calls run under zero-data-retention terms. Every extraction is logged and traceable to its source document.

OCR and document parsing; a small open-weight model in-environment for PII masking; Claude (Sonnet-class) under zero-retention API terms for extraction and validation reasoning; REST integration into the firm's practice-management system; human-in-the-loop review queue.

Re-keying eliminated for standard submissions; intake turnaround from days to same-day; errors caught at the door instead of found downstream.

Before attorneys could evaluate a case, someone had to read medical records arriving as hundred-page scans — days of senior time per case. The records contain protected health information, and a summary an attorney can't verify line-by-line is worse than no summary at all.

A pipeline that digitizes incoming records, builds a chronological medical timeline, links every statement to its source page, and drafts a case summary a paralegal verifies before it's used.

Records never leave the firm's own cloud tenancy: models are consumed through the cloud provider's hosted endpoints with a business associate agreement in place, so PHI stays inside one compliance boundary. Zero data retention on model calls; access mirrors the firm's existing case permissions; every summary sentence cites its source page.

OCR tuned for medical scans; embedding-based retrieval over each record set; Claude via Amazon Bedrock inside the firm's AWS account for citation-grounded summarization; an evaluation harness that measures citation accuracy before any prompt or model change ships.

First case evaluation in hours instead of days — with every claim traceable to the page it came from.

Leads arrived by phone, web form, and marketplaces — and sat unanswered during storm season while every crew was on a roof. Slow response meant lost jobs.

One queue for every lead, automatic answers to routine questions, structured collection of job details (address, photos, roof type), estimates drafted for the owner's approval, and follow-up that continues until the job is scheduled or closed.

Lighter here, but real: consumer contact data is minimized and consent-tracked for text follow-ups; call recordings are transcribed, then discarded on schedule.

Voice and form capture with speech-to-text; Claude (Sonnet-class) classifies leads, answers routine questions, and drafts estimates for the owner's approval; CRM and calendar integration via webhooks.

Response time from “when someone got to it” to minutes; quotes out same-day; no lead falls through the cracks.

Dozens of programs, with policy knowledge living in binders and long-tenured staff; onboarding slow, grant reporting heavy. The organization works with children and families — so deciding what the assistant would never see was the first engineering decision.

A private assistant that answers staff questions from the organization's own program documentation and policies — every answer citing its source document — plus drafting support for recurring grant reports.

The knowledge base holds policies, program manuals, and reporting templates — deliberately no case records and no personal data about families. Access rides the organization's existing single sign-on, scoped per role. Answers cite a source document or decline.

Retrieval over a curated document store; Sonnet-class model for cited Q&A and report drafting with staff in the loop; runs inside the organization's existing cloud workspace.

New staff self-serve routine answers; program leads get hours back; report drafts start at eighty percent instead of zero.

A retailer of heavy, time-sensitive goods ran its storefront, delivery scheduling, and carrier coordination in disconnected tools. Exceptions — a wrong address, a weather hold — were handled by phone and spreadsheet.

A pipeline connecting storefront orders, delivery-slot scheduling, and routing, with automated exception handling and clean human handoff for the cases that need judgment.

Customer addresses handled with minimal retention; payment data never enters the system — it stays with the payment processor, where it belongs.

Deterministic orchestration first: storefront and routing APIs, queues, retries, audit trail. An LLM only where judgment helps — triaging exceptions and drafting customer notifications. One dashboard as the ops team's single screen.

The ops team runs the day from one screen; exceptions get handled proactively instead of after the truck left.

An agency assembled client reports by hand every month from ad platforms, search data, and analytics — senior time spent on copy-paste, delivered too late to act on.

A data pipeline consolidating every platform per client, and an AI layer that drafts a weekly performance brief — what moved, why, what to change — reviewed by the account lead before it goes out.

Aggregate marketing performance data only, segregated per client workspace; no personal customer data enters the pipeline.

API connectors (ad platforms, search console, GA4) feeding a small warehouse; scheduled transforms; the weekly brief drafted by a frontier model from the numbers — never the other way around — with account-lead review before send.

Reporting went from a monthly scramble to a weekly rhythm; account leads spend the saved time on decisions.

Leadership knew AI mattered but not where it made financial sense — and vendors were pitching everything. Resident data raised real compliance exposure that most pitches ignored.

A structured evaluation of their operations: where AI genuinely helps, where plain software wins, where neither is worth the cost — including a data-governance review. Then a working prototype of the highest-value case: a family-facing assistant answering from the organization's own published materials.

The phase-one prototype was deliberately grounded only in public and marketing materials — zero resident data — so it could ship while data governance matured on its own timeline.

Department-by-department assessment; prototype built on retrieval over published materials with a frontier model and guardrails; feedback capture wired in to inform the phase-two decision.

A roadmap with verdicts and honest costs instead of a slide deck; the prototype shipped as phase one.