I spent 15 years on the buying end of enterprise SaaS. Most of it was too big, too slow, and built for everyone instead of for the work in front of my operators. Modern agentic tools changed the equation, so I went back to writing the code myself.

Applied AI implementation leader. Operator-builder. 3x founder, 4x exit operator. Named inventor on patent PCT/US2016/053346.

I run Brightline Labs, an applied AI software studio building lightweight, fit-for-purpose tools for the operational workflows enterprise SaaS does not handle well.

• 6 products shipped in 5 months across production scheduling, AI job search, agent auditability, task orchestration, decision-support, and open-source utilities
• 11 years on npm — recollie first published in 2015, modernized to v2.0.0 in 2026 (the gap is the point)
• 4 exits across 4 distinct models — services-firm sale, $20M IPO on CSE, PE platform acquisition, mid-market rollup
• 112 repositories total, 7 worth featuring (below)

• Applied AI implementation
• Enterprise workflow automation
• Production scheduling and inventory planning
• Open-source LLM agent observability, replay, and auditability
• Domain-specific internal tools for operators
• Agentic development workflows with independent verification

The most commercially useful vertical products stay private; the public repos are methodology, infrastructure, companion apps, and open-source utilities — now led by LLM Workbench, open-sourced under MIT and published to npm. The pattern is deliberate: ship the domain-specific products private, open-source the infrastructure around them.

A 90-second walkthrough of the LLM relevance scoring and feedback loop:

The useful pattern in my background is not "executive learns to code."

I wrote JavaScript before the current AI wave, published an npm package in 2015, and came back to software because agentic tooling changed what one domain expert can ship. The real edge is that I have lived inside the workflows I now automate — production planning, purchasing, compliance, inventory control, sales operations, executive reporting, and cross-functional operating cadence.

My best software ideas usually start with a sentence I heard, said, or muttered as an operator:

There has to be a lighter way to do this.

Before Brightline, I led or founded four businesses to exit across four different models:

• Carbon Chemistry (managing director, 2022–2023) — specialty-chemicals turnaround; sold to Linde via mid-market rollup
• Flor Americas (founder & MD, 2019–2022) — biotech innovation consultancy across 7 countries, $0 → $10M+ peak revenue; sold to a Miami-based PE platform
• PLUS Products (co-founder & COO, 2015–2018) — regulated CPG manufacturer from concept to IPO on the CSE (CSE:PLUS); $20M raise at ~$100M valuation; named inventor on PCT/US2016/053346
• Brightacre Law (founder, 2010–2014) — professional services firm scaled to 16 attorneys across 3 offices; sold to a national firm

I have run P&Ls, scaled multi-state and multi-country manufacturing, led facility buildouts, implemented ERP / CRM / MRP, raised capital, executed acquisitions, and worked through the operating reality that software is supposed to help.

• Domain model before interface polish
• Typed data contracts and explicit validation at every boundary
• Observability for LLM calls and agent behavior, not vibes
• Human-reviewable AI workflows; independent verification before merge
• Lightweight where it can be, framework-free where it should be
• Tools operators can trust on a Monday morning at 7 a.m.

• Languages: TypeScript · JavaScript · Python
• Frontend: Next.js (App Router) · React · HTML / CSS
• Backend & data: Node.js · Prisma · Drizzle · Postgres · MySQL/TiDB · Redis/Upstash
• AI / agents: OpenAI API · MCP (Model Context Protocol) · open-source LLM control plane (LLM Workbench) · custom feedback-learning loops · three-LLM advisor/builder/verifier pipeline
• Infra & delivery: Vercel · Vercel Cron · Resend · GitHub Actions
• Tooling: Cursor · Codex · Claude Code

• Production scheduling built from scratch in Next.js
• Open-sourcing an LLM control plane and making it genuinely installable
• The 11-year gap between recollie v1 and v2
• Why "lightweight and fit-for-purpose" beats "configurable and comprehensive" almost every time
• Running a three-LLM advisor / builder / verifier pipeline as a real software-delivery process

• LinkedIn — linkedin.com/in/roymcfarland
• npm — npmjs.com/~roymcfarland
• Email — roy@brightline.io