Complex Systems Physicist turned Data Scientist. Founder @data4sci, Fellow @isi_fondazione, Formerly @jpmorgan, @NYUDataScience, @trmlabs
Manhattan, NY
Joined June 2008
- Tweets18K
- Following501
- Followers3.8K
- Likes1.3K
Running this live tomorrow at 12pm ET — one real support ticket, a dozen questions in one pass. Bring your edge cases.
Tomorrow, 12pm ET: Jev in One Call. 30 min, free, live on Zoom.
Meet Choice, Score, Noul on a real support ticket — a dozen questions in one pass.
Register: maven.com/p/1d62f8/jev-in-on…
I keep seeing teams treat agent safety as a dependency to install rather than a property to engineer. Happy to be told I'm wrong — that's what the replies are for.
If you need a framework to make your AI agent "safe," your agent isn't safe.
Safety is engineered into the system. Not imported from a package.
Curious what others think — am I wrong?
(Building this live on Oct 31 → luma.com/qwhit9op)
I'm teaching this one live — a real support ticket, the 3 Jev primitives, one ~100ms call. Free, Oct 21. Would love to see you there.
Free lesson — Wed Oct 21, 12pm ET.
Meet the 3 primitives of Jev on a real ticket. One ~100ms call, $0.042/1M tokens, output free.
Register: maven.com/p/1d62f8/jev-in-on…
The model is the easy part — the system around it is where most agents fail. I'm covering the full engineering layer (permissions, memory, supervision) in my Oct 31 workshop.
Really looking forward to this one — four hours of hands-on harness engineering, building everything that separates an agent demo from one you can actually ship. Join me on Oct 31! 🎃
I'm running a live 4-hour workshop on Oct 31: Build an AI Data Analyst you can actually trust — MCP tools, guardrails, context management, fault tolerance. Laptop-only, framework-free Python. 🎃
Details + tickets: luma.com/qwhit9op
The agent failure I lose sleep over: numbers that look right and were never computed. This is the guardrail that catches them - and on Oct 31 we build the whole harness, live.
Last day to grab a spot — tomorrow I'm showing how to wire an LLM eval gate into GitHub Actions so bad merges block themselves. Free, 30 minutes, notebook included.
Last call — free LLM eval gate session is TOMORROW (Oct 7, 2 PM EDT).
Regression thresholds + GitHub Actions wiring + merge blocking, all in 30 min.
Register: maven.com/p/07c4b3/stop-bad-…
The model is rarely the problem - the missing harness is. That's what my Oct 31 workshop is about: permissioned tools, guardrails, recovery, in plain Python.
After years of putting agents into production, this is the lesson that stuck: the model is the easy part. On Oct 31 I'm teaching the hard part.
An AI agent isn't reliable because the model is good.
It's reliable because of the harness: permissioned tools, context budgets, guardrails, and recovery.
The model suggests. The harness decides.
Building that harness, framework-free, Oct 31:
luma.com/51oqj4cn
The model was never the hard part. If your agents are headed for production, this thread is the checklist I wish I'd had a year ago - and on Oct 31 I'm teaching the whole thing hands-on.
17 days to go. We build the whole harness live — calibrated judges, error bars, and a CI gate that blocks bad merges. Come build it with me.
17 days until the LLM Eval Harness workshop.
Build a production eval harness live in 4 hours. Calibrated judges, error bars, CI gates.
Engineers who measure ship faster.
maven.com/bruno-goncalves/ev…
I've watched too many teams discover a bad prompt change from user complaints instead of CI. A gate in the merge path is the boring fix that works.
This is the setup I wish every team I consult for already had: evals in CI, regressions blocked before they merge. Teaching it free on Oct 7 — come build it with me.
A bad prompt change ships the same way a good one does: nobody measured either.
Free session Oct 7 — I'll show you how to wire an eval gate into GitHub Actions so regressions die in the queue, not in prod.
maven.com/p/07c4b3/stop-bad-…
Last call — I'm teaching this free session today at 2 PM EDT. If your eval scores jump around and you can't tell signal from noise, this hour is for you.
Last chance — free LLM eval session is TODAY at 2 PM EDT.
Learn to put error bars on your scores in 20 lines of Python.
Stop chasing phantom regressions. Start trusting your numbers.
maven.com/p/48c82e/put-error…
I've modeled epidemics for years — this time I pointed the same machinery at our own LLM habits. The trap-door result surprised even me: dependence that outlives the conditions that created it.
One month out! If you've been meaning to get serious about evals, this is the one — we build the whole harness live, from calibrated judges to CI gates.
4 weeks until the LLM Eval Harness workshop (Oct 16).
Calibrated judges. Bootstrap CIs. CI gates. 4 hours. You build it live.
Seats are filling.
maven.com/bruno-goncalves/ev…
I've lost count of the eval dashboards I've seen with zero error bars. This is the fix — and I'm teaching it free on Sep 23.
My favorite part of building this one: you write the model as a string and epidemik does the rest — R0, diagram, YAML. Give 0.2.0 a spin.
Define an epidemic model as a string:
"S + I -> I + I", beta=0.2 → compute R0, draw the diagram, save YAML.
epidemik 0.2.0: ODE compartmental models, networkx graph spread, metapopulation travel matrix — one MIT-licensed pip package.
pypi.org/project/epidemik/
Flaky eval scores cost me more debugging hours than actual bugs ever did. This is the method I use to tell real regressions from noise — teaching it free on Sep 23.