Ranga Sampath | Network Expert | AI Things Builder

Building Agentic AI tools to solve high stakes problems in Cloud Networking.

The Network Ghost Agent encodes senior architect methodology into safe-exec code.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

I went dark for three years. The podcast downloads didn't stop.

Post January 2022, I stopped the You+AI podcast — it had reached Top 8 globally in its category. Full guest pipeline. Sent them all declines. Later, took an enterprise engineering role and went heads-down.

Didn't delete anything.

💡 Eight months after stopping: still getting monthly downloads.
Twelve months. Twenty months.
Requests for new episodes and guest recommendations arriving in 2023, 2024, even into 2026.

In 2020 I had written about four things I believed made innovation work:

*Be Curious. Embrace Change. Ride the Wave. Partner.*

🔍 What I found when I came back in 2026 — to build open-source agentic AI tools for cloud network troubleshooting — was that all four had kept running without me.

My website archive did the work I was not doing. The timestamped work log outlasted any portfolio I could have built. The name You+AI turned out to describe a thesis, not a topic: human expertise and AI capability working together, whoever the human is, whatever the domain.

The tools I am building now are the most honest expression of that thesis yet.

Dormant is not dead. The foundation is real, or it isn't. Stopping is not the same as failing.

Original 2020 recipe, built to last, read the full recipe @ https://t.co/4ygvyDAxat↗

#PersonalBrand #Innovation

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@hnshah Signed up

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

Yes this is what I have realised over time ie it is best to keep context separate for unrelated tasks. I use Claude projects to separate the tasks. Each project gets its own knowledge base.

Related tasks can be different conversations in the same project if they did indeed share the knowledge base of files uploaded to the project.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

Recently I experienced the non-determinism first hand when upgrading models.

What used to work for the customer no longer worked-- the underlying behaviour of model had changed. The behavioural contract of how models will work for your use case in your domain isn't spelled out anywhere.

Wrote about this -- https://t.co/kaVgUTs8wD

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

LLM Vendors tell you about 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝘄𝗶𝗻𝗱𝗼𝘄 𝘀𝗶𝘇𝗲𝘀 and 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝘀𝗰𝗼𝗿𝗲𝘀

They cannot tell you what the model actually does on your infrastructure.
..
..

That gap is the 𝘣𝘦𝘩𝘢𝘷𝘪𝘰𝘳𝘢𝘭 𝘤𝘰𝘯𝘵𝘳𝘢𝘤𝘵 — and it's invisible until you deploy.

📊 I ran 30 real network investigations (10 uses cases * 3 models) — Gemini 2.5 Flash, Sonnet 4.6, Claude Haiku 4.5 — with identical system prompts and fault scenarios. All three identified the root cause correctly. But here's what did not work out as expected:

🔍 𝗥𝗲𝗺𝗲𝗱𝗶𝗮𝘁𝗶𝗼𝗻 𝗱𝗲𝗽𝘁𝗵: Does the model understand what the fix touches beyond the immediate problem? One model knows that `--service-endpoints` replaces the entire list and includes existing endpoints in the command. The other two omit them. *SQL connectivity fails silently.*

🔧 𝗧𝗼𝗼𝗹-𝘁𝗼-𝗵𝘆𝗽𝗼𝘁𝗵𝗲𝘀𝗶𝘀 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆: When the model forms a hypothesis about a specific layer — OS firewall vs. traffic shaping vs. routing — does it pick the right inspection tool? Or does it use a sender-side packet capture to verify a routing blackhole (spoiler: wrong tool, wrong answer)?

⚠️ 𝗕𝗮𝘀𝗲𝗹𝗶𝗻𝗲 𝘂𝘀𝗮𝗴𝗲: If you provide reference data in the prompt, does the model use it? Or does it investigate from scratch and take 10x longer?

🛡️ 𝗙𝗶𝘅 𝘀𝗮𝗳𝗲𝘁𝘆: Same diagnosis. Same confidence: high. Entirely different blast radius depending on whether the model understands what the command does to adjacent config.

They properties don't appear until you run the model on your actual domain, against your actual use cases.

💡 I documented what each model does well, where it struggles, and how to build your own evaluation framework before you migrate production traffic. Because when the vendor deprecates a model, the migration guide covers API compatibility. It doesn't cover whether the replacement still knows to verify at the effective route table instead of a PCAP.

The behavioral contract you validated is valid for the model you validated it on. 𝗪𝗵𝗲𝗻 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗰𝗵𝗮𝗻𝗴𝗲𝘀, 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗿𝗮𝗰𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀.

Read the full details @ https://t.co/jejh93xsba↗

#AIBuilders #LLMModels

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@ShivaJmc12 Just spotted the female 2 days back in Wayanad

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@isvictoriousss While Wordpress continues on...

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@quxiaoyin What will be your Top 3 tips for someone willing to switch to DeepSeek?

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@GergelyOrosz Build a StackOverflow over LLM returned responses in the org 😜

Show the top 10 repeated queries over the hour in a dashboard visible to everyone.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@GergelyOrosz Show engineers how much of their token budget is used up rather than meeting them know only after their budget is exhausted.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@schmittdev @aronkor Thank you for the pointer.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

Thank you for sharing @dharmesh -- insightful piece.

One problem that I've faced recently which related well with this article is what happens when you migrate to a newer and shinier model.

Things break and you need to solve it for the customer if you are building off the Yellow Brick Road. Wrote about my experience here -- https://t.co/kaVgUTs8wD

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@joeschmidtiv Thank you for putting this together. Will have to read a few more times to internalise.

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

A specific problem that I can relate to with this article -- LLM model changes and the push by vendors to upgrade to the latest and shiniest model. Turns out you need to testiest and re-validate a lot of what worked before when the model changes.

This is what you need to do for the customer when you are building off the Yellow Brick Road. Wrote about it here -- https://t.co/kaVgUTs8wD

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@Aiswarya_Sankar What would be some strategies to best use token towards production features ie how could we minimise wastage?

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

I swapped the AI model in a production network forensics agent.

The confidence scores didn't change. The fix recommendations did.

🔍 30 runs. 10 fault scenarios. Same prompt. Same Azure infrastructure. Three models:
Gemini 2.5 Flash, Sonnet 4.6, Haiku 4.5.

-- Overall diagnosis accuracy: 90%.
- -Remediation safety: 73%.

Every unsafe fix — delivered at confidence: high.

⚠️ The clearest example: a missing service endpoint. All three models found it. Two gave
a fix that would silently delete another service endpoint added during that morning's
maintenance window. One gave the safe command — and explicitly warned that the CLI
argument is a full list replacement, not an append.

Same diagnosis. Same confidence score. Entirely different blast radius.

💡 That's the behavioral contract — the implicit agreement between your system prompt
and a specific model about how to reason and what a fix does to adjacent config.

That contract is NOT documented by any vendor.

When the model changes, the contract changes. The migration notice won't mention it.

I wrote this up in full — 3 failure modes, what held cleanly across all three models,
and what a remediation safety eval actually looks like.

Check it out ➡️ https://t.co/kaVgUTrAH5

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@saybwala Seriously valuable cheat sheet

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@captsanthoshkc How may we source this?

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@ParveenKaswan So beautiful!

[Ranga Sampath | Network Expert | AI Things Builder @youplusai]

@supriyasahuias @MegamalaiS @tnforestdept @UNBiodiversity @UNEP @WWFINDIA @WWF @IUCN @dhanu_paran @shanmughanandam @pnsenthilkumar @sriram_p_murali To state "once existed" breaks my heart 💔 We must do better.