Problem. Approach. Outcome.

CS-01 · 2025 — ongoing

Verin: audit-ready AI for life sciences

Verin — compliance intelligence platform · with affiliate Genisage LLC

The problem

Life-sciences quality teams sit on thousands of controlled documents — SOPs, CAPAs, deviations, training records — and the tools that could read them are exactly the ones a quality unit can't trust: generic AI chat that answers without citations, ignores PII, and leaves no audit trail. Meanwhile compliance status is scattered across spreadsheets, and the question “are we inspection-ready?” takes a week to answer.

The approach

So we built Verin — a compliance-intelligence platform for pharmaceuticals, biologics, and medical device organizations, developed with our life-sciences affiliate Genisage. Verin reads your document corpus and answers questions with citations back to the source documents: every response shows the work behind it, and drafting the next document starts from evidence rather than a blank page.

Under the hood: AI-powered document classification, PII/PHI redaction before content reaches a model, semantic search across the corpus, and a real-time dashboard across six quality domains — data integrity and 21 CFR Part 11, CAPA effectiveness, deviation management, FDA quality metrics, training compliance, and inspection readiness — with Part 11-aligned audit trails throughout. Private by design.

The outcome

Verin is live: a trusted compliance advisor that reads your documents, drafts the next one, and shows the work behind every answer — the audit-ready-by-construction pattern we bring to every engagement, shipped as a product. verin.app tells the full story.

CS-02 · 2026

An AI assistant that never invents

Principal consultant's public practice site

The problem

A consulting principal wanted his site to answer visitor questions about his background and services around the clock. The obvious answer — bolt on a chatbot — carried an unacceptable risk for a professional reputation: generic assistants confidently invent facts, and a chatbot that misstates your credentials is worse than no chatbot at all.

The approach

We built the assistant around a single curated knowledge profile that drives both the site's visible content and the assistant's system prompt — one source of truth, edited in one place. The assistant streams answers from that profile and is instructed, tested, and constrained to decline anything the profile doesn't cover rather than guess.

Every visitor question and answer is forwarded to the owner by email, turning the assistant into a lead-capture and follow-up channel rather than a black box. The underlying model is configurable by environment variable, so the practice can adopt better models as they ship — without touching code.

The outcome

The site now gives accurate first-contact answers at any hour, the owner reads every conversation, and updating a single profile file updates the résumé page and the assistant's knowledge together. No invented facts — by architecture, not by hope.

CS-03 · 2023–24

Guardrails for enterprise generative AI

Add Value Machine Inc — enterprise AI security platform, Austin TX

The problem

In 2023, enterprises were racing to put generative AI in employees' hands — and every prompt was a potential data leak. Sensitive content flowing to third-party models, no audit trail of who asked what, and security teams forced to choose between blocking AI entirely or watching shadow usage grow ungoverned.

The approach

Our principal worked with Add Value Machine on their enterprise platform that sits between a workforce and the foundation models: prompt policies that catch sensitive data before it leaves, role-based access controls, audit logging of every interaction, and routing across AI providers so no single vendor becomes a blind spot.

He brought the discipline of a decade of digital-assistant engineering — evaluation, measurement, and quality regression — to a product whose whole job is making AI behavior provable rather than assumed.

The outcome

The pattern this work proved out is the one we now bring to every AI engagement: employees get the models, security gets the controls, and compliance gets the evidence — adoption without abdication.

CS-04 · 2026

Retiring a WordPress estate

General Internet Group portfolio — gigviz.com, seancochrane.us & affiliates

The problem

Three WordPress sites had accumulated across the portfolio, each on its own aging AWS instance — Lightsail here, EC2 there — each needing PHP patches, plugin updates, and backups nobody enjoyed verifying. DNS was spread across Route 53 and GoDaddy, email authentication was incomplete, and the monthly hosting spend bought mostly maintenance risk.

We run our own properties the way we advise clients to run theirs, so the portfolio became the engagement.

The approach

Every site's content and media was archived first — nothing depends on the old servers existing. Each property was then rebuilt as a static-first Next.js application on Vercel, with contact forms moved to a transactional email API (spam-trapped, no plugins) and analytics replaced with a lightweight privacy-respecting layer.

DNS was consolidated to Cloudflare with mail-preserving cutovers: Google Workspace MX records untouched, apex and www records switched cleanly, and legacy records carried over deliberately rather than dropped. Email authentication was completed across the portfolio — SPF, DKIM, and DMARC brought to full pass with alignment, and DMARC set on a path to enforcement.

Then the old estate was actually retired: instances terminated, elastic IPs released, hosted zones deleted, and both old server IPs verified dark.

The outcome

The portfolio now deploys by git push in under two minutes, has no servers to patch, and costs almost nothing to host. Every property is faster than its WordPress predecessor, mail authenticates everywhere, and the AWS accounts bill zero.