May 28, 2026 · Adeo / Decathlon Dev Summit
Bullsh*t AI
is over.
Real analytics with
just a prompt.
Mehdi Ouazza · MotherDuck
md · adeo-decathlon
You've all seen this before.
md · adeo-decathlon
Who I am
Mehdi Ouazza aka mehdio.
Developer Advocate at MotherDuck · DuckDB & analytics infra.
Data infrastructure for ~10 years — Axa, Klarna, Back Market, Trade Republic.
Sharing knowledge & teaching data people for 4 years — YouTube, blogs. 1M+ views, still counting.
Burned by AI demos too. So today, I'll be honest.
md · adeo-decathlon
First question
What actually happened in the last 6 months?
md · adeo-decathlon
Well, for me :
md · adeo-decathlon
The real signal
Even the skeptics flipped.
md · adeo-decathlon
Skills · sharing & attention
One markdown file. 159k stars.
A single
CLAUDE.md out-stars most frameworks. People share skills like code — because it's just markdown.md · adeo-decathlon
Demo 1 · 2 minutes
Old model vs new model.
Same prompt. Same harness. The old one guesses. The new one fetches, reads, runs, fixes — the agentic loop, live.
opencode · OpenRouter · DuckDB
md · adeo-decathlon
SQL is solved.
Python pipeline scaffolding — a prompt away.
Connecting a model to your warehouse — MCP made it trivial.
From "do you know SQL?" to "do you know what should be asked?"
md · adeo-decathlon
The engine
DuckDB — and why agents love it.
10M
downloads / week on PyPI
- › Local — runs where the data is, no network
- › In-process — a library, not a server
- › Zero infra — perfect agent target
+ scale-up
- › Infra for the answer
- › Multiplayer — share live
- › Scale beyond your laptop
- › AI-native — bring your own agent
md · adeo-decathlon
Demo
MotherDuck MCP.
Plain English in. SQL + result out. No UI in between.
Claude · MCP · MotherDuck
md · adeo-decathlon
"Can I write this?" → "Do I know what and how should I build this?"
Complex projects: meh. Models don't refactor a 12-domain warehouse alone.
Hallucinations: way down. Not gone. Guardrails = skills, MCPs, evals.
Combine models: Opus + Codex gets me through bigger refactors. One model drifts.
Humans: lower technical barrier → educate more people with less data background.
md · adeo-decathlon
Everyone is empowered for analytics.
Tech expert · non-tech alike
Jordan Tigani · MotherDuck CEO — "a problem that took me the better part of a day, asked Claude to do the same thing."
md · adeo-decathlon
Data engineering is AI engineering.
Parallel agents on branches hitting checkpoints = wave dispatch = a DAG.
"How in the loop do I stay?" = batch vs streaming = state.
The scaffolding that made DE hard? AI drafts it. You review and own it.
md · adeo-decathlon
Demo
Dive.
English → dashboard.
No dragging. No chart picker. The user shapes it after the fact.
MotherDuck Dive · interactive
md · adeo-decathlon
What is Dive
Bring your own agent. Get an interactive data app.
User
·
BYA · bring your own agent
→
prompt
"create a chart of weekly signups"
→
Dive
Sandbox · JS
const chart = vega.spec({
data: motherduck.query(
"SELECT week, COUNT(*) AS signups
FROM events GROUP BY 1"
),
mark: 'bar',
});
hosted on
The agent writes the chart. The sandbox runs the JavaScript. The data never leaves MotherDuck.
md · adeo-decathlon
Agents need SaaS.
Neon: 4× more databases created by agents than humans in 7 months. The SaaS that survives is the one agents can call — malleable, API-first, sandbox-friendly.
md · adeo-decathlon
The user brings the AI. The platform sandboxes it.
User
↓
asks via any agent
·
python
java
golang
javascript
→
Sandbox · the platform
Internal
service
service
SaaS
tool
tool
task executed
Reviewer
↑
reviews via any agent
·
←
Same controlled platform. Any agent on either side — one builds, one validates. Human in the loop on each end.
md · adeo-decathlon
Demo 4 · 3 minutes
Flights.
Vibe-coded ingestion.
From "I need this daily" to a running scheduled job in minutes. Pipelines nobody loved writing.
MotherDuck Flights · vibe-code · schedule · monitor
md · adeo-decathlon
The platform of the future
One prompt. A team of agents. The SaaS orchestrates them.
User
·
"I need the sales from all countries — past 2 years"
→
SaaS · agent orchestration
Check agent
data already there?
→ query · or no?
→ query · or no?
↓ NO
→
Flights agent
create pipeline
ingest sales data
ingest sales data
→ DATA READY
→
Dive agent
dive on this domain?
no → build a new one
no → build a new one
→ DASHBOARD
The SaaS isn't a single product anymore — it's a team of specialist agents the user's agent calls. Each one does one thing well.
md · adeo-decathlon
Monday morning — what to do.
N° 01
Pick a SaaS you use — BI, pipelines, notebook — and build an AI MVP of it.
N° 02
Set a token budget. $$$
N° 03
Stop watching AI demos. Experiment.
md · adeo-decathlon
md · adeo-decathlon
Q1 · last 6 months
Q2 · for analytics
Q3 · what the future looks like