Brian Baldock

10 Followers
5 Following
92 Posts
PronounsHe/Him
A “working” AI Workbench project isn’t the same as a usable one.
Part 2 of my DGX Spark series covers the path from a validated container to a real training environment, including why Jupyter and package management silently failed. 👉 https://blog.brianbaldock.net/learning-fine-tuning-on-nvidia-dgx-spark-part-2
Before you can fine-tune anything on DGX Spark, you have to get the platform right. Part 1 of my blog post walks through AI Workbench, PyTorch, 👇 CUDA, and the things that break early on Blackwell. 👉 https://blog.brianbaldock.net/learning-fine-tuning-on-nvidia-dgx-spark-part-1

It’s been about a month since I published the LightningCopilot guide, integrating Microsoft Copilot Studio Agents into Salesforce Lightning with Entra ID SSO and MSAL.

What’s interesting is how many teams I've talked to saying they hit the same blockers:
🔹 Entra ID redirect mismatches
🔹 Power Platform API confusion
🔹 CSP and Locker issues in Salesforce

The takeaway? These integrations aren’t just technical, they’re architectural!

The hardest part of the LightningCopilot project wasn’t building the LWC.
It was getting Entra ID, Power Platform, Copilot Studio and Salesforce to all agree on who you are SILENTLY.

Federated credentials, custom scopes, Static Resources instead of CDNs… a dozen small things that had to line up.

If you’re working on similar integrations or just curious how to make Copilot Studio live securely inside Salesforce, the full blog is up now.

#CopilotStudio #Salesforce #EntraID #PowerPlatform #SSO

AI agents aren’t hype, they’re the next step in automation.

81% of orgs expect them in production inside 18 months. If you want to see how to build your own (and not just talk about it), the Microsoft Copilot & AI Agents Summit runs Nov 26–27.

Learn from real-world builds and go from concept to production.
#MS-FTE #Microsoft #Copilot #Agents

Drift often happens when agents read their own wrong outputs. Limiting exposure fixes it. Recent studies shows context management-like sliding windows helps models sustain longer horizons. In my work, I externalize state to trusted systems and feed the bare minimum back in. Give it a shot!
#ThursdayThoughts #AI #LLM #AIAgents
Voting across multiple outputs looks comforting, but in long tasks it adds little. The research shows sequential compute, ie: “thinking tokens”, is what prevents drift. Use reasoning where errors matter most; you’ll get more reliability per token than from brute force voting.
#ThursdayThoughts #AI #LLM #AIAgents

Identity & access management made practical.
Check out Microsoft Entra deployment videos—hybrid join, MFA, Conditional Access & more.

Real-world guidance for IT admins:

My desk looks like a small tornado passed through, coffee mug, cables, notebooks, and a few rogue spinner magnets that didn’t survive another fidget session.
Some people need calm to focus; I apparently need creative chaos.
Curious, what’s the one item on your desk you can’t live without?
#DeskSetup #FidgetThings #OrganizedChaos #WorkVibes