BioCreative · Commercial Operating System

Commercial growth, engineered.

Every commercial operation has to do the same things — research and segment a market, build inbound and outbound pipelines, run a sales process and a customer journey, and field the team and processes to carry it all the way through closing.

Most companies stitch that together from a CRM, data vendors, a rotating cast of consultants, and internal sales and marketing hires.

We engineer it into one operating system — a spine, a brain, a memory, and a directed agent fleet — with a small, super-empowered team in the loop, working it firsthand to drive real results. It’s grounded in proven commercial practice built over more than a decade in the field.

This is the methodology and the architecture behind it.

Propose → a human confirms → dispatch → work ships → the system learns

The OS in seven layers

  • FoundationBrand bible spine, modular context (Karpathy), ICP & market research, Miller–Heiman.
  • Brain & memoryRAG + Conductor + perpetual self-learning memory.
  • Fleet & harnessesA directed agent workforce, spun up many ways.
  • Commercial enginePipeline agent, SSO follow-up, meeting reply bot, per-stage content.
  • Marketing channelsWebsite, SEO, GEO, paid, social — one shared graph.
  • Build & evolveBuild-along delivery, skill catalog, ownership ladder.

The thesis

Software is written once and runs forever. An operating system is continually optimized.

That reframe is the whole point. Karpathy’s framing is our north star — the LLM is the new processor, context is the program, and tools, files, and memory are the rest of the computer. But an operating system is never finished: it rides the exponential curve of AI, so as new models launch the system improves — and it keeps improving on its own.

Processor

The LLM does the thinking

Frontier and local models, routed by cost and task — the cheapest model that can do the job, with the expensive ones reserved for genuinely hard reasoning.

Program

Context is the software

A modular context system aligned to each project, agent, or harness — the right files, rules, and memory loaded for the job, nothing more.

Tools

What it can do

Skills, workflows, and agents — the packaged capabilities the processor reaches for to actually get work done.

Files

What it knows

SQL databases and the knowledge base — the structured truth the system reads from and writes back to.

Memory

How it learns

Keeps track of everything that happens and is the backbone that lets the operating system self-learn and sharpen over time.

The architecture

A four-part body, not a pile of tools.

We model the company as a living body, not a stack of modules. Each part has one job — and together they run in a continuous loop, which is exactly why the system stays legible as it grows.

Brain

Reasoning & memory

Where knowledge lives and decisions get made — retrieval, routing, and judgment.

Heart

Circulation

The steady pulse — schedules, triggers, and the plumbing that keeps everything flowing, with no thinking of its own.

Nervous system

The operating system itself

The knowledge graph and persistent memory that route signal through the body and back to the brain — the substrate that ties heart and brain into one self-learning loop.

Eyes

How we see & steer

The interfaces a person uses to watch the system and direct it — the human’s window into the whole thing.

The closed loop

The Conductor proposes. A human confirms. Nothing runs on autopilot.

Each workday the Conductor reads the Heart’s pulse, reasons over the Brain, and writes proposals into Cortex — the operator surface (Now · Inbox · Pipeline · Plan · Dispatch). The operator confirms, defers, edits, or drops. Approved work dispatches several ways — visual workflows, declarative pipelines, a long-running task, an edge function, or a multi-model council — and the outcome feeds memory. The same loop that runs BioCreative is the loop we stand up for the companies we work with.

SenseTranscripts, emails, signals, and market data flow in and resolve to accounts & opportunities.
ProposeThe Conductor (a CEO-agent) drafts the next best action each workday — a follow-up, a piece of content, a scoring pass.
ConfirmIt lands in the Cortex Inbox. A person confirms, defers, edits, or drops it — the confirmation discipline is enforced.
DispatchApproved work runs the right way for the job: a visual workflow, a declarative pipeline, a long-running task, an edge function, or a multi-model council.
LearnOutcomes write back to memory; a weekly consolidation pass distills them. The system gets sharper every cycle.

The honest boundary: the closing, the signatures, and the relationship stay human. The OS removes the drag so the people can do the parts only people can do.

Where it already runs

Not a concept deck. A system in production.

This isn’t a roadmap of things we might build. We run the most advanced version of this operating system on BioCreative itself — the same brain, fleet, and closed loop described here — and we stand up tailored versions of it, build-along and in the open, for the companies we work with. The strongest proof is that the machine markets and sells using itself.

Take the tour

Seven layers, one machine.

01

Foundation

The immutable spine — brand bible, modular context, ICP research, sales methodology. Open →

02

Brain & Memory

RAG, the Conductor, and a memory that compounds. Open →

03

Fleet & Harnesses

A directed agent workforce and the many ways we spin agents up. Open →

04

Commercial Engine

Pipeline agent, SSO follow-up, meeting reply bot, per-stage content. Open →

05

Marketing Channels

Website, SEO, GEO, paid, social on one shared graph. Open →

06

Build & Evolve

How we deliver, anonymize skills, and hand over ownership. Open →