Work.

Founder had a hypothesis: outreach triggered by life-event signals (new entity filings, real estate transactions, cross-state relocations) would outperform generic intent-based prospecting for tax and accounting services. Needed validation with real data inside a single client instance before investing in multi-tenant infrastructure.

• •AWS-serverless architecture: Lambda, Step Functions, EventBridge, DynamoDB, API Gateway, fully deployed via CDK
• •Signal pipelines for new entity filings (Cobalt Intelligence API + Florida Sunbiz parser + custom Virginia SCC scraper), real estate deeds (ATTOM Data with 133 Virginia jurisdictions), and cross-state relocation detection via buyer-previous-address cross-reference
• •Enrichment Step Function with Clay, Apollo, Hunter.io, plus email verification, so no unverified addresses reach outreach
• •Rule-based scoring with deduplication, suppression list, and configurable outreach-hold scheduling
• •Deliverables: architecture doc, build log, runbook, provider setup guide, daily output to HighLevel

Hypothesis testing pipeline running end-to-end on real Virginia + Florida filings and deeds, ready to validate response rates before any multi-tenant investment.

Chief Administration Officer holding the quality, consistency, and risk line across 10-plus centers and four service lines. The question wasn't whether AI was promising. It was whether AI could give 10 centers the consistency of one without adding headcount and without putting the families they serve at risk.

• •8 to 12 stakeholder working sessions across HR, IT, Quality, and three site directors across service lines
• •Site visits at two to three centers across early childhood and housing
• •Findings synthesized against four criteria: impact, feasibility, risk, time-to-value
• •Explicit "will not recommend yet" list: no client-facing AI, no AI in mental health delivery, no vendor recommendations
• •Six-question board-ready Opportunity Map plus a separate one-page Risk & Governance Brief for board, funders, and accreditation bodies
• •90-day sequenced roadmap, lowest-risk-highest-leverage first

Board-ready document delivered with a 90-minute live readout. Top three plays identified with named owners. Carmen left the engagement with a defensible answer to her board on what AI the org is and is not doing.

SAP Business One was the gravity center of the business. Nearly every operational pain point across finance, sales, client experience, and operations traced back to one of three SAP interaction patterns: getting data out, getting data in, or querying it. CFO had a 15-item automation priority list. The team needed a clear-eyed mapping of where AI belonged versus where deterministic automation was the right answer.

• •Five departmental audit interviews across Sales/CX, Finance, and Operations Leadership (including the CEO)
• •CFO's 15-item automation spreadsheet cross-referenced against transcripts to find universal pain points
• •Pain points mapped explicitly to AI Production Framework pillars: P1 (Design Doctrine), P9 (Economic Architecture), P10 (Operational Sovereignty)
• •Honest delineation between LLM tasks and tasks that should be scraping + templating + rules (the LLM Elimination Checklist applied)
• •Quantified ROI per pain point using real volume estimates (80+ daily SAP extractions in Finance alone)

Implementation roadmap with named pain points (PP-01 through PP-N), framework-mapped ownership, and a clear sequence that respects the team's bandwidth. The CFO's own "matching, accessing SAP and downloading, formatting" theme validated and prioritized.

Mature agency with 25 years of domain expertise, enterprise clients, and a $1M internal platform already built. Not looking for a 90-day transformation pitch. Looking for someone who could get on the field, score quickly, and build toward something durable. The destination they were quietly working toward: a proprietary capability that encoded two decades of institutional knowledge as a competitive moat.

• •Phase 1: G Suite maximization. Gems for opportunity briefs, proposals, SOWs, creative briefs, meeting recaps, project handoffs, pre-mortems, milestone reviews. Notebook LM knowledge bases for delivery process and per-client context.
• •Drive governance: standardized folder structure mapped to the 5-phase delivery process, naming conventions, ownership rules so the library stays organized as it grows
• •Three live training sessions, run personally, not delegated. Recorded for future onboarding.
• •Phase 2 plan: HubSpot, Teamwork, and Slack integration, scoped separately
• •Phase 3 plan: proprietary MCP server encoding the creative process, client context, and methodology as a moat

Phase 1 designed with Phase 3's destination in mind, so nothing built in months 1-2 has to be undone when the agency builds its proprietary layer. Senior judgment plus junior execution velocity in the same engagement.

Solo founder had already done the hardest part. 27 jurisdiction rules documents. One converted into a 13-page AI-ready reference with if/else logic. A Filipino VA with two years of construction permitting experience. A live Zoho CRM. A working cold email campaign. He was past the "figure it out" stage. He needed a tightly-scoped 6-week build that proved one jurisdiction worked end-to-end. A competitor had pitched LangChain + GPT-4 + generic Playwright at $1 to $2 per workflow.

• •AWS Lambda plus event-driven messaging instead of LangChain. Same orchestration benefits, no demoware fragility at scale.
• •AWS Bedrock for inference with model evaluation up front (not pre-selected). Fraction of GPT-4's token cost, validated via benchmark on real permits.
• •RAG + scripted chain-of-thought: the AI walks the founder's existing if/else jurisdiction tree step-by-step and shows its work, so the VA sees exactly where any wrong call happened
• •Doer/evaluator multi-agent pattern: one agent fills the permit, a second agent checks the first against the same rules doc before the VA sees anything
• •Tauri desktop app for the non-technical VA (free, included)
• •Enterprise serverless Playwright platform provisioned for Phase 3 portal submission (also included, separately valued at $30K+)

One jurisdiction working end-to-end. A playbook to onboard jurisdiction #2 without me. A reference benchmark on the engagement: a prior transcription framework I built priced at $0.0003 per transcript versus competitors at $0.30. 1000x cheaper, same quality. That's the cost discipline this build inherits.

Team building a QA tool that operates on MindSmith JSON and markdown, not rendered web pages. Industry research says ~30% of WCAG is fully automatable and ~10% partially automatable on rendered pages. Their surface was narrower (content-only) and they needed an honest, defensible scoping document before writing any code.

• •Mapped all 230 best practices from the reference WCAG library against what's actually checkable from content JSON
• •Three-bucket architecture: deterministic code-based checks (no AI), AI-assisted checks (LLM reasoning required, every flag is a recommendation a human approves), out-of-scope checks (require browser, assistive tech, or human interaction)
• •Bucket 1: 31 code-based checks. Alt text, heading hierarchy, table headers, dead links, language tags.
• •Bucket 2: 23 AI-assisted checks. Meaningful alt text, descriptive link text, color-not-the-sole-indicator, sensory instructions, heading descriptiveness.
• •Bucket 3: explicit documentation of the 176 best practices the tool cannot legitimately check, and why, so the team doesn't oversell coverage
• •V1 coverage stated honestly: 54 best practices (23.5%) across 21 categories

A scoping document that gates V1 development with a defensible coverage claim, an evaluation harness target (the 80% accuracy threshold for Bucket 2), and a clear separation between what the tool guarantees and what it recommends. The kind of pre-build artifact that prevents "we said it did X" lawsuits later.