I help founders and leaders build AI products that earn trust, scale responsibly, and keep humans in the loop.

AI, which I call Synthetic Agency (SA), is a great technology in the right context, is a harmful technology in the wrong context, and is a low ROI feature in most contexts.

I've built synthetic agency in products from the ground up, launched ventures in the field, performed academic research and contributed to think tanks, wrote The Ethics of AI, and advised 100+ leaders over the past 12 years.
I’ve learned that, while most people attribute failure modes to technicalities or talent, the true failure is philosophical.

The products that succeeded long term had teams who did not skip the hard questions. They built synthetic agency when they had to, not because they just could. They solved real problems by listening to their customers without chasing trends.

If you're working on AI, with AI or are figuring out how it impacts your work, and want to:
• Build something people actually want to use 
• Think through the ethical stuff without getting paralyzed
• Understand how your brand discovery and promotion should change with AI engines in town

I might be able to help.

I also write for 15,000 people on LinkedIn and for 3,000 on my newsletter about honest tips to make sense of this technological shift.

I'm also launching a monthly private newsletter. Invite-only. For Founders, CMOs, and Heads of Growth who need to control their brand voice in AI conversations (before it becomes too late or too expensive). DM me if you want to get an invite.

🎤 Open to speaking, podcasts, and collaborations—DM me or visit albertochierici.com

Alberto Chierici

I advise and help companies do more with their data. Solutions Engineering @ Snowflake | Mentoring | Storytelling | Data all day everyday

I'm Doneyli De Jesus, and I've spent over 20 years proving that the most successful AI initiatives occur when technical expertise meets strategic leadership and flawless execution.

I translate C-suite goals into reality by uncovering bottlenecks, aligning stakeholders, and driving change management that unlocks AI-driven competitive advantage. 

What sets me apart is the rare combination of strategic thinking, leadership, and execution that turns AI vision into measurable business impact.

My approach bridges the gap between boardroom strategy and operational excellence:

Strategic Diagnosis – I lead comprehensive AI strategy audits that map workflows, quantify pain points, and deliver executive-ready roadmaps ranked by impact and effort, with clear ROI prioritization.

Rapid Solution Design – When pain points are known but paths aren't clear, I orchestrate focused discovery sprints combining stakeholder interviews, rapid prototyping, and implementation planning to move from insight to action in weeks.

Execution Leadership – I don't just strategize—I deliver. Whether leading technical teams through workflow automation or rolling up my sleeves with Python and SQL, I ship solutions that create significant value.

Change Management Excellence – I design adoption programs that transform pilots into permanent competitive advantages, complete with training frameworks and internal champion development.

This track record of consistent delivery makes me a tremendous asset to organizations seeking AI transformation with real-world impact.

I also contribute to the broader AI leadership community through selective speaking engagements and strategic advisory work.

Most importantly, I'm a husband, a proud father, and someone who believes authentic leadership creates lasting organizational change.

Note: All opinions are my own and do not reflect those of my employer.

Doneyli De Jesus

You can't improve what you can't measure
Like in Supply Chain, our agentic production line needs quality control and monitoring

Most companies deploying AI agents right now share the same blind spot:

> They can't reproduce the output. <

If you can't trace what an agent did, why it did it, and whether it would do the same thing again; you can't improve it. 

You can't audit it. 
You can't trust it.

We solved this in manufacturing a century ago. 
Every production line has measurement, verification, and documented procedures. 
AI agents need the same rigor.

The good news: we have the infrastructure now.

OpenTelemetry gives us standardized tracing. 
Tools like LangChain and LlamaIndex integrate with observability stacks. 
And databases like ClickHouse can ingest and query billions of traces in real time.

This is the work I'm doing at ClickHouse right now, helping teams instrument their AI systems so they can actually see what's happening inside.

I talked about this with New Tab by Island, exploring what agent observability looks like in 2026 and why it's becoming non-negotiable for production deployments.

♻️ Share with someone who needs observability 
💬 Are you tracking your agents? Let me know in the comments!

Solutions Architecture @ ClickHouse | AI Architect & Strategist | TEDx & Public Speaker

I'm at the Ski hill with the family ⛷️ 
While Claude does some work work work! 

I have a "Family Office" Wealth Advisor
To have holistic view into our family's finances, future, planning, etc.

I've been meaning to migrate the database to ClickHouse (of course 😉)

I as the architect take care of the following: 

1. Design very well defined and structured plans

2. Scaffold the user journeys and outcomes than I'm looking for

3. Validate that AI recommendations about data model, storage decisions, etc. Are within my expectations and best practices

4. Make sure to have discrete units of work derived from the main plan / spec 


Then I add Claude to each individual feature (github issue) and let it go to work.

Once I'm back home in a couple of hours I can review the work, do user testing to make sure is working according to my specs and give the final sign off.

This is how AI is meant to be used. The human ideates, designs and architects. AI executes while you do the fun stuff in your life.

AI enabled Work-life balance 🤗

♻️ Share this with someone who can benefit 

Who else is doing this?

💬 Let me know in the comments or DM to chat about it.

20 years working with databases.
With agentic workflows, the differences are exponential.
The slower your engine, the more it compounds.

I covered that and more at our ClickHouse NYC Meetup => https://lnkd.in/eQxeN9-n

Here's the core insight:

> Your database (not your LLM) is likely the bottleneck in your AI strategy.

The "Agentic Math" Problem:

When a human asks a question, we expect ONE response. 
But when an Agent reasons through a prompt It might run 5 to 50 queries to stitch together an answer.

If your database takes 1 second per query:
50 queries = 50 seconds. (User experience: Broken).

If your database provides sub-second latency:
50 queries = 10 seconds. (User experience: Meaningful).

Two shifts I'm seeing in production-ready AI:

1. Stop building "God Agents."
Don't build one agent that knows everything. 
You'll hit context window limits and invite hallucinations. 
Build discrete specialists with skills. 
In my demo, I showed how a Hotel specialist and a Maintenance specialist can coexist, providing higher-quality insights by staying within their domain.

2. Moving from "What" to "Why."
Basic SQL can tell you what your sales were. 
Agentic reasoning tells you WHY a specific hotel is outperforming the rest. 
By using MCP (Model Context Protocol), we're abstracting the database nuances so the LLM can focus on reasoning, not just retrieval.

The practitioner's reality: AI is at its best when it augments humans, but that only works if the data backbone can keep up with the speed of thought.

If you're still waiting for your analytics to finish before you can finish your coffee, your agents are already falling behind.

#ClickHouse #AIAgents #DataArchitecture #RealTimeAnalytics

Meta just acquired Manus, the AI agent startup with $125M ARR.

Jack Dorsey called it "excellent." Some called it "China's second DeepSeek moment."

But here's what most coverage is skipping:

→ Leaked code suggested Manus is essentially a wrapper around DeepSeek's API with browser automation
→ Independent tests showed 40% accuracy drops vs. their published benchmarks
→ Their social accounts were suspended over crypto scam concerns
→ Kai-Fu Lee publicly called for "rigorous, open testing"

So why did Meta buy it?

Because they're not buying the tech. 
They're buying the traction.

$125M ARR. Real users. A product people actually pay for monthly.

Meta can swap in their own models later. What they can't easily build is the distribution, the UX that got people paying, and the team that shipped it.

This is the playbook for AI agent acquisitions in 2025:

Big Tech isn't looking for the best model.
They're looking for the best product-market fit.

If you're building agents, stop optimizing for benchmarks.

Start optimizing for "will someone pay for this?"

I just refactored 12 AI agent skills.
Here's what changed.

Before:
→ 200-400 lines per skill file
→ Everything documented upfront
→ Dense, intimidating to read
→ Claude had to load everything just to understand the basics

After:
→ 20 lines per skill file
→ Checklist-driven workflows
→ Documentation on-demand
→ Progressive disclosure pattern

The difference? Cognitive load.

When you're building AI agents (personal or enterprise), your instinct is to document everything. Every edge case. Every option. Every template.

But that's not how people work. Or how AI works, for that matter.

The pattern I used:

1. Checklist first — 5-6 steps, checkbox format
2. References second — Links to detailed docs when needed
3. Dense content externalized — Moved to PATTERNS.md, CONTENT_TYPES.md

Result: 1,156 lines removed. Same functionality. Faster execution.

This isn't just about Claude Code skills. It's about how you architect any system where AI needs to make decisions.

Your RAG system? Don't dump the entire knowledge base into context.
Your agent workflows? Don't preload every possible path.
Your prompts? Don't front-load every instruction.

Progressive disclosure works because both humans and AI systems perform better when they:
• Start with the minimum needed to begin
• Pull in details as required
• Don't carry cognitive overhead for unused paths

Look at your longest prompt, workflow, or config file. 
Can you break it into "start here" + "details when needed"?
💬  Let me know in the comments if you've had dealt with this

💾  Save this pattern.
♻️  Share with others if helpful

🍁 Toronto ClickHouse user group 🇨🇦 
Join us on Feb 19th
for our 1st meetup of 2026 👇 

https://lnkd.in/eKpKS4kK

This will be an evening of learning, food, and great conversations with the ClickHouse community.

​​Hear from database and systems experts as they share real-world stories of building high-performance analytics and trading systems.

​​Whether you’re a data engineer, backend developer, or systems enthusiast, this meetup is a great chance to learn, ask questions, and connect with others in the community. 

RSVP to secure your spot, and we look forward to seeing you there!

I've been a Claude Code power user for months now.
Terminal commands. MCP servers. Custom agents. 
The whole stack.

So when Anthropic launched Cowork this week, 
I saw it immediately:
Same foundation. Different door.

Here's what the launch signals:
✅ Cowork uses the exact same Claude Agent SDK as Claude Code. 
✅ Same model. Same capabilities. 
❗ But instead of the terminal, you just point Claude at a folder and chat.

That's not a product launch. 
That's a market expansion signal.

The pattern I'm watching:
> Oct 2025: Claude Code gets a web interface
> Dec 2025: Slack integration launches
> Jan 2026: Cowork eliminates the terminal entirely

Three months. Three UX layers. Same underlying power.

Why this matters:
1. Developer tools are becoming mainstream tools; not by dumbing them down, but by building new interfaces on top of them.
2. The technical capability stays constant. The accessibility layer keeps widening.

This is happening everywhere:
> Coding assistants → business automation
> Data pipelines → natural language queries
> Infrastructure as code → infrastructure as conversation

IDC predicts 80% of enterprise workplace apps will have AI copilots embedded by end of 2026.

The infrastructure is set. 
The UX race has begun.

For technical leaders, this creates a window:
> The teams who understand both the capability AND the interface will shape how organizations adopt these tools.
> Translation still matters; it's just shifting from "explain AI to executives" to "design how AI shows up in workflows."

Cowork isn't the future. 
It's the signal that the future is arriving faster than most orgs are ready for.

Are you building the capability or the interface? 
Because both matter now.

45 minutes of review. 
5 hours of work (by claude)
Full day skiing with my daughters.

Be present with family OR be productive. 
Pick one.
AI broke that trade-off. 
But only if you structure it right.

Here's the Architect-Executor framework I'm using
And I was present with my daughters.
That's not "working remotely." 
That's a different relationship with work.

Steal the Full breakdown + spec template => https://lnkd.in/eYbGdZ5R

#AI #SoftwareEngineering #WorkLifeBalance #WorkLifeIntegration

Open article: Get AI to Work While You Enjoy Your Life! by doneyli.substack.com

Get AI to Work While You Enjoy Your Life!

89% of executives say AI is a priority.
And yet, many AI initiatives never make it to production.

You think AI takes 12-18 months. It doesn't, 30 days for a production MVP.

You think you need ML PhDs. You don't; you need product discipline.

You think you need a perfect strategy first. You don't; shipping reveals strategy.

Most companies treat AI like a bolt-on feature instead of rebuilding how they build.

The fix? A prioritization matrix.

Plot every use case on two axes: Feasibility × Impact.

Four quadrants:

- Quick Wins (high impact, high feasibility) — Start here
- Strategic Bets (high impact, hard to build) — Plan for later
- Incremental (low impact, easy) — Delegate
- Kill (low impact, hard) — Kill now

The teams I see succeed pick ONE Quick Win and ship it in 30 days.

Not 30 AI initiatives. Not a 12-month roadmap.

One use case. One metric. 30 days.

Hype doesn't pay the bills. Shipped models that move KPIs do.

♻️ If you’re picking AI use cases right now, save this 💾 
💬 What's your Quick Win for Q1?

A VP of Engineering messaged me:
“94% eval accuracy in staging. Shipped Friday.
 Monday morning: support is drowning.
 
AI is correct… but useless. Cloud bill doubled.”
“What did we miss?”

Most teams evaluate LLMs like traditional software.
That’s the mistake.

Traditional QA is mostly pass/fail.
 LLM performance is continuous and multi-dimensional.

Here’s the framework I use in production:
— Performance (latency, timeouts, reliability)
— Quality (helpful and grounded, not just “correct”)
— Cost (tokens, retries, infra burn)

Teams that succeed in production aren’t the ones with perfect staging accuracy.

They’re the ones instrumenting all 3 dimensions continuously with:
 — confidence signals
 — user feedback capture
 — regression evals tied to real tickets

Miss one dimension, and production will find it for you.

I shared the full framework + a 4-week implementation checklist 👇 

♻️ Repost to your network for others 
💬 Drop a comment. Which dimension blindsides your team most often? Performance, quality, or cost?

3 Dimensions of LLM Evals

Multiple AI models debating each other 
just hit 85% accuracy on complex diagnoses.
Experienced physicians scored 20% on the same cases.

This is Microsoft's MAI-DxO and it's built on what Mustafa Suleyman calls "Chain of Debate."

Here's why it matters:
> Chain of Thought was chapter one. 
A single model thinking out loud, showing its reasoning from A to B.

> Chain of Debate is chapter two. 
Multiple models: GPT, Claude, Gemini, Llama, Grok; acting as a virtual panel of specialists. 

Debating. Cross-checking. 
Building consensus.

Suleyman calls it "the AI version of peer review."

The system creates five AI agents, each with distinct roles: 
1. Generating hypotheses, 
2. selecting tests, 
3. challenging assumptions. 

They negotiate in real time before reaching a diagnosis.

Important context:
This was a proof-of-concept, not a production deployment. 

The benchmark used deliberately difficult "stump the doctor" cases from the New England Journal of Medicine, not routine diagnoses.

And the physicians in the study worked without access to colleagues, textbooks, or AI tools. In reality, doctors rarely diagnose complex cases alone.

What this signals for enterprise AI:
1. The future may not be one model to rule them all.
2. It's specialized models in conversation, catching each other's blind spots, and showing their reasoning transparently.

That visibility into how AI agents influence each other could be key to building trust in high-stakes decisions.

Still early. 
But worth watching closely.

taplio

The Battle

Alberto Chierici

Doneyli De Jesus

Followers

Posts

Engagements

Likes

Performance ScoreDoneyli De Jesus's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alberto Chierici

Doneyli De Jesus

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Doneyli De Jesus's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreDoneyli De Jesus's LinkedIn Performance