Louis-Paul Baril

🎯 Consultant IA spécialisé en solutions locales et sécurisées. J'aide les entreprises à adopter l'intelligence artificielle sans compromettre leurs données.

[Louis-Paul Baril @LPBaril]

The weirdly hard part of memory is not storing more, it is deciding what should be safe enough to surface before a meeting.

[Louis-Paul Baril @LPBaril]

The hardest part of shipping coaching features is resisting the temptation to turn them into surveillance.

[Louis-Paul Baril @LPBaril]

Trustworthy company memory has an annoying requirement: it has to know when it might be getting old.

[Louis-Paul Baril @LPBaril]

Today BusinessOS got a little closer to turning messy conversations into accountable work instead of another pile of notes.

[Louis-Paul Baril @LPBaril]

The weirdest part of building trustworthy AI is that rejection needs product design too.

[Louis-Paul Baril @LPBaril]

I spent today on the kind of work nobody screenshots, but every serious customer eventually asks about.

[Louis-Paul Baril @LPBaril]

Today’s tension was privacy versus usefulness: BusinessOS needs to remember that "Robert likes Bordeaux" without ever turning someone’s personal life into loose company data.

[Louis-Paul Baril @LPBaril]

I thought cookie consent would be a small legal checkbox, but NEXT-186 turned into a real architecture lesson.

[Louis-Paul Baril @LPBaril]

I keep bumping into the same tension with BusinessOS: the Brain needs to be powerful enough for outside agents to use, but boringly strict enough that I can sleep at night.

[Louis-Paul Baril @LPBaril]

I shipped a feature yesterday that feels small in UI and huge in philosophy.

NEXT-196 adds buying-signal classification to BusinessOS.

[Louis-Paul Baril @LPBaril]

Yesterday I merged PR #465 for BusinessOS: sensitivity, lifecycle_status, and authority labels across the Brain-retrievable tables. Plus a re-stamp CLI that walks old memory provenance and only promotes rows when the mapping is unambiguous.

[Louis-Paul Baril @LPBaril]

Data sovereignty is mostly boring code.

[Louis-Paul Baril @LPBaril]

I keep making the same mistake.

PR #447 landed yesterday — a real-time job hook that lets agents pause mid-workflow and wait for human input before resuming. Clean spec, clean code, merged.

PR #456 landed right after — fixing everything the spec got wrong.

[Louis-Paul Baril @LPBaril]

I just shipped 437 lines of code that could break everything.

But I'm weirdly proud of the part that stops hackers.

[Louis-Paul Baril @LPBaril]

For businesses, the takeaway isn't "switch tools tomorrow." It's that the AI landscape is bigger than the chatbot you're using. Different problems need different math under the hood.

Full video here (Turing Post does great work): https://t.co/RcclQzmseO

[Louis-Paul Baril @LPBaril]

Quick primer on something most business owners haven't heard of yet but should:

Energy-Based Models (EBMs)

https://t.co/RcclQzmseO

#AI #MachineLearning #EnergyBasedModels

[Louis-Paul Baril @LPBaril]

Yann LeCun (Meta's chief AI scientist) has been arguing for years that this is the path forward — world models, JEPA, energy-based reasoning. Less autocomplete, more actual understanding of constraints.

[Louis-Paul Baril @LPBaril]

This sprint: trajectory analytics, a personal development coach, pattern learning from top performers, structured table retrieval. NEXT-166 through NEXT-169, all landed.

Then I spent my "day off" fixing CSS.