Introducing HumanDB: The World's First Human-Powered Analytical Database

The Case for Artisanal Data

Think about what happened to bread. For decades, we industrialized it. We made it faster, cheaper, more shelf-stable. We optimized the hell out of bread. And then, right when we perfected the factory loaf, people started paying $9 for a sourdough boule from a guy named Søren who ferments his own starter passed down from the Oregon Trail and only bakes on Tuesdays.

AI is doing to data work what factories did to bread. We are rapidly approaching a world where every query, every dashboard, every insight can be generated instantly by a machine. And it'll be good. It'll be really good. We've seen it ourselves — our MCP server routinely finds things in our data that we didn't even know were there.

But that's exactly why human-crafted data is about to become a luxury. When everyone has access to instant, machine-generated analytics, what becomes scarce? The human touch. The analyst who pauses before answering. Who squints at the number. Who says, "Well, technically the query returns 4,287, but I was here when we onboarded that batch of test accounts, so the real number is probably closer to 4,100."

That's not a bug. That's artisanal. That's hand-selected, small-batch, locally-sourced insight. And with HumanDB, you can build your entire data stack on it. When they're at their desk. Which is most of the time. Usually.

The Impedance Mismatch No One Talks About

There's a dirty secret in analytics: most of the time you spend "doing data work" isn't spent querying. It's spent figuring out what to query. Understanding context. Knowing that the revenue_final_v3_ACTUAL table is the one you actually want, not revenue_final_v3 or — God forbid — revenue_final. Remembering that Q3 numbers look weird because someone changed the fiscal calendar in 2019 and nobody updated the docs.

You know who remembers all of that? Dave. Dave has been here since before the codebase. Dave was here during the 2016 migration. Dave knows that the employee_status field has seven possible values, but only three of them mean anything, and one of them — "active (legacy)" — is a lie.

No amount of dbt documentation is going to capture what Dave knows. We tried. Dave's knowledge is not structured. It is not in a catalog. It is stored in a combination of muscle memory, sticky notes, and a spreadsheet on his desktop called FINAL_USE_THIS_ONE.xlsx.

So we asked ourselves: instead of trying to get all of Dave's knowledge into the database, what if we just made Dave the database?

Architecture: Blazingly Slow by Design

HumanDB uses what we're calling Innovative Single-Tenant Architecture™. Here's how it works:

1. You send a query via our Python client.
2. The query is forwarded to Dave's phone as an SMS.
3. Dave does the math.
4. Dave records a voice memo with the answer.
5. You receive dave_answer.mp3 as well as a speech-to-text transcript.

That's it. No query optimizer, because Dave optimizes based on vibes. No cache invalidation, because Dave just remembers. No cold starts, though Dave does take a minute to get going before his first coffee.

We should be upfront about availability. HumanDB offers what we're calling Presence-Based Compute. The system is live when Dave is at his desk, which our monitoring shows is roughly 73% of business hours. The other 27% is lunch, bathroom breaks, and that thing where Dave walks to the kitchen, forgets why he went there, and ends up talking to someone from marketing for twenty minutes. We're working on it. Dave is working on it. He just needs to grab a coffee first.

If you're used to DuckDB's sub-second query times, this will be an adjustment. Our benchmarks show HumanDB query latency at a very competitive 2–4 business hours. Or 3–5 if it's quarter-end. Dave has a lot on his plate.

OLAH: The Processing Model the Industry Has Been Waiting For

OLAP. OLTP. We've had these acronyms for decades, and honestly, has anyone's life gotten better? HumanDB introduces OLAH — OnLine Analytical Humans — a processing model that has been industry-standard since 2003 and is powered entirely by drip coffee and determination.

OLAH has several advantages over traditional architectures. First, there's no cold storage problem, because nothing is stored cold. Everything is in Dave's hippocampus, which runs at a comfortable 98.6°F at all times. Second, OLAH provides built-in gut-feel analytics. When Dave says "that number looks off," it usually is. No statistical test required — just decades of institutional knowledge and a vague sense of unease.

Third, and perhaps most importantly, OLAH supports both SQL and natural language. Dave learned SQL first, then English. He understands both. You can write a perfectly formed query, or you can walk up to his desk and say "hey, how are we doing?" and he'll know you mean revenue.

Eventual Consistency (Dave Will Get Back to You)

Some databases promise strong consistency. Others promise eventual consistency. HumanDB promises emotional consistency. Dave is pretty sure the data is right. Like, 80% sure. He'll circle back if he realizes he was wrong, which, to be fair, is more than most dashboards do.

We've also implemented what we call Post-it Indexing. It's sub-optimal, but colorful. Each index is handwritten and stuck to the monitor bezel. When Dave goes on PTO, we photograph the monitor and email it to the team. It's our version of a backup strategy.

The SLA is one business day. Unless it's quarter-end, in which case it's three. Unless Dave is on PTO, in which case it's five. Unless it's quarter-end and Dave is on PTO, in which case you should probably just wait for Dave.

Dave vs. The Machines

We ran some benchmarks. Dave graded himself, which we think is fair.

The numbers speak for themselves. Or rather, Dave speaks for the numbers, because that's his job.

Try It Yourself

HumanDB is available today. You can install it with a single command:

Copy code
pip install humandb

Or, if you prefer the one-liner:

Copy code
uv run --with humandb python -c "import humandb; print(humandb.query('how are we doing'))"

Dave doesn't actually have access to your data. He makes it all up. That's the point.

We also have a live demo on our website, where you can query Dave in real time. Try asking him SELECT count(*) FROM orders or, if you're feeling existential, can we migrate this to Snowflake?

Pricing: Simple, Human

We offer three tiers:

The Honest Truth

Look, we've spent the last year showing that AI can genuinely replace a lot of the manual work in analytics. Our MCP server answers questions about our business that would have taken a human analyst hours to figure out. Dives let AI build interactive data apps in minutes. The robots are, objectively, very good at this now.

But in a world where machine-generated insight is abundant and instant, human judgment becomes the rare ingredient. The artisanal stuff. The thing you can't scale, can't automate, and can't reproduce — because it lives in the head of someone who's been staring at your data longer than your company has had a Slack workspace.

Every now and then, you need someone who just knows. Someone who remembers the migration. Someone who can look at a number and say, "that's wrong, I can feel it." Someone who will message you on Slack at 11pm to say "hey, I was thinking about that query you ran, and I think the join was off."

Dave is that person. And today, Dave is a database. When he's at his desk.

Happy April 1st. If you want to see what MotherDuck's actual AI-powered analytics can do, check out our MCP server and Dives. They're faster than Dave. But Dave has better vibes. And he remembered your birthday.

Built with ❤️ and mild existential dread by MotherDuck.