The unit of change is a workflow, not a chatbot.
Delightful Computer is Morgan. Production AI for workflows that already have a reviewer, a record, and a cost of being wrong. The offer is a production system for one named workflow — mapped, constrained, shipped, instrumented, and handed over with the code, runbook, and evaluation suite.
Proof: Eli Lilly regulated-lab audit intelligence
For Eli Lilly, Delightful Computer mapped an FDA-regulated laboratory data flow and deployed a reasoning-model pipeline that clusters logs, flags out-of-spec events, and explains root-cause in plain English. The system integrates with existing laboratory systems under strict controls. The client-reported result is >$2M/year in operational savings, delivered in an 8-week deployment.
Senior scientists had been spending 40+ hours a week reconstructing audit trails from LIMS exports, ELN notebooks, and instrument logs. Out-of-spec events surfaced weeks after the fact. Any automation that touched GxP data had to ship with validated change-control, electronic signatures, and immutable logs under 21 CFR Part 11. The architecture clustered deviations statistically first and used the language model second, so the system degrades to a deterministic tool when the model is uncertain. Validation artifacts (IQ/OQ/PQ, risk assessment, traceability matrix) were written in parallel with the code. Every write action is Part 11-signed; every model call is logged with prompt, response, and model version.
Reported results from the case: audit prep time from months to days; >$2M annually in operational efficiency; 12 critical anomalies caught in the first three months that manual review missed; scientists reclaimed for analysis instead of data compilation. The number was reported by a Director of Quality Assurance at Eli Lilly.
Also in high-trust work
A second regulated, high-trust case is the multi-agent legal research copilot at Vorys Sater Seymour: 50× cost reduction per research matter, 6-week pilot to production. Role-specialized agents handle statutes, case precedents, and jurisdictional filters, with citation verification before anything returns to a partner. The firm now licenses the tool.
Other production systems live on the work index — including financial-services pursuit briefs, construction parts-data automation, upstream lease-action intelligence, and a law-firm prototype suite. Home-services and side-project work is not ranked with the regulated-lab proof.
Who
Delightful Computer is Morgan, working solo. Engineering sits on the critical path from workflow map to deployed system: the same person writes the code, sits with the subject-matter experts, and owns the outcome. Roughly a decade designing production software for regulated enterprises, the last several years on LLM-native systems.
Good-fit work has a measurable operational bottleneck, subject-matter experts who know the job better than any model, constrained data access, a human reviewer with real authority, and a clear definition of what production means. In practice that has meant regulated and high-trust settings most often — the constraint set, not the vertical, is what makes the work tractable.
Services
Work takes one of three shapes. Discovery answers what should we build: a workflow map, data-access plan, architecture spike, and a go / no-go recommendation. Production Spike answers will it work in production: one workflow shipped end to end with an evaluation suite, guardrails, observability, alerting, and runbook transfer. Partnership answers will it still work next year: continuous eval monitoring, the model migration path, a bounded monthly change budget, named-hours incident response, and a quarterly written report. Partnership is the one engagement that starts cold, including a system inherited from another vendor with nobody on staff who can maintain it.
Engagements are fixed-scope and quoted against one named workflow rather than a public rate card. The edge is explicit: what workflow ships, what data it can touch, who reviews outputs, what the model may decide, what it logs, and what happens when it is wrong.
Writing
Field notes on agent architecture, evaluation-driven development, RAG quality, telemetry, structured outputs, workflow routers, and durable agent operating systems. Written for technical buyers and builders who need implementation language rather than generic AI commentary.
Contact
The fastest path is the calendar: book 20 minutes. A written note is fine if a slot does not work. Send it via Contact or morgan@delightful.computer. Do not send confidential records, secrets, regulated samples, patient data, or production credentials until an access pattern is agreed. NDA on request.
Security, privacy, and developers
Delightful Computer designs for client-owned deployment, scoped credentials, least-data access, reviewer-visible reasoning, logging, and rollback. The public website does not expose customer workspaces, operational write actions, or a public production API. Client data handling is defined per engagement through the statement of work, data-access plan, security review, and target architecture. Security reports: morgan@delightful.computer.
Public notes: Security, Privacy, Developers. Machine-readable files: llms.txt, llms-full.txt, well-known llms.txt, agent skills index, sitemap.xml.