Juliana McMillan-Wilhoit
F&T Labs · AI Operations · Ongoing

Sometimes the pain point is mine. The AI stack I run my own company on.

Our company's marketing department is me. Our research assistant is me. So I engineer for that the same way I engineer for clients: find the real bottleneck, weigh cost against security, and decide, on purpose, whether to buy the tool or build it. This page is the part of my work that never ends, my ongoing practice.

Flourish & Thrive Labs AI Architecture Build vs. Buy Automation Security
Impact
2-person
Company that ships video, content, and research like it has a studio
$0 ads
Behind a 1,000+ mailing list, built organically
Every one
Integration security-audited by me before it touches our work

The user I interviewed was me. I was drowning in the same twenty decisions.

Editing a talking-head video is the same twenty decisions, every time. Years of voice memos, the thinking behind every project, sat unsearchable on a phone. And every promising AI tool that could help came with the same unanswered questions: where does the data go, what does it cost at scale, and who checked?

A two-person company can't hire its way out of any of that. It has to engineer its way out.

The build-vs-buy rule: if the work is deterministic, build.

I almost bought a video tool. I'd assumed the value was in the editing itself. Then I noticed that most of editing isn't creative at all. It's deterministic. Cut the silences, sync the captions to the transcript, apply the brand treatment. If I can write down the decisions, I build it as code I own. That's the rule I actually use: if the workflow's decisions can be written down, build the pipeline; if the value is someone else's ongoing R&D, buy it, after the security review.

And the inverse discipline matters as much: knowing what not to automate. Some steps stay manual on purpose, because the judgment is the value, or because the failure mode of automating it is worse than the labor of doing it.

What runs every week

A transcript-driven video pipeline: edits, silence-trimming, word-synced captions, brand treatments, and fully animated product demos, all in code. One recording in, a finished reel out.
A voice-memo research database: every memo transcribed and embedded into a queryable vector database, with the cost-versus-security trade-offs weighed deliberately at each step. Years of out-loud thinking, now searchable. The research behind this site ran on it.
Security audits on every integration: I review what each tool can access and where data flows before adopting it, instead of asking the AI to vouch for itself. In government-adjacent work there's a wrinkle most vendors miss: AI conversations can be public records. I architect for that.
A deliberate not-list, the work that stays human: the writing that is the relationship, the analysis I need to be able to interrogate line by line, the judgment calls that are the actual job. I hand AI the parts of the job that don't need my judgment, so I have more time for the parts that do.
Impact
A two-person company that operates like it has a studio, a research team, and a security office.

The pipeline ships polished video without a production hire. The research database turns years of voice memos into an asset instead of an archive. The audit habit means every tool in the stack was adopted with eyes open: cost, security, and public-records exposure priced in.

When I recommend an AI architecture to a client, it's the same reasoning I use on my own company every week.

Stack & Practices
Python Pipelines Speech-to-Text Vector Database / RAG FFmpeg Security Review Build-vs-Buy Analysis Cost Modeling
Next project
No marketing budget. So I built the growth engine myself, and I still run it.
Growth Engine