Sulay Joshi

Full Stack .NET Engineer & Solution Architect. C#, Azure & AI — enterprise apps across web, desktop & mobile. Exploring Semantic Kernel + MAF. Follow along!👇

[Sulay Joshi @sulayjoshi]

@VisualStudio Goodbye static class boilerplate and fake instance methods. True extension members are exactly what we need to clean up domain interfaces and express intent without the old syntax workarounds. Absolute gold.

[Sulay Joshi @sulayjoshi]

One bad response with no "regenerate" button = a user who stops trusting the feature entirely.

Cheap correction isn't polish. It's the difference between recoverable and broken.

#DotNetAI #ProductDesign

[Sulay Joshi @sulayjoshi]

AI feature UX checklist:
✅ stream, don't spin
✅ show uncertainty
✅ let users edit/regenerate
✅ label AI content clearly

Thread on why this matters more than model quality 👇

[Sulay Joshi @sulayjoshi]

A spinner + wall of text feels slower than streaming, even at the same total latency.

Perceived speed ≠ actual speed. Users judge the former.

#AIUX

[Sulay Joshi @sulayjoshi]

@dotnetfdn This is exactly where multi-agent orchestration needs to go. Shifting from a single generalist prompt to a runtime network of persistent specialists—with distinct roles,charters,and standup ceremonies—is how we move from basic coding assistants to true digital engineering squads.

[Sulay Joshi @sulayjoshi]

@VisualStudio Treating model outputs as plain text is the fastest way to break a production app. Moving toward strictly structured formats, rigid schema enforcement, and tool-calling validation is the only way to build predictable, reliable software with LLMs.

[Sulay Joshi @sulayjoshi]

@dotnet Chatbots wait, but automated agents execute. Moving from reactive text inputs to fully managed schedules, event listeners, and self-reminders means we can finally build truly autonomous systems instead of just better wrappers. This GA is a massive win!

[Sulay Joshi @sulayjoshi]

User closes the tab? Cancellation token should stop the model call, not just the response.

Handle OperationCanceledException as a clean exit, not an error.

#AspNetCore #CSharp

[Sulay Joshi @sulayjoshi]

Users don't wait 8 seconds for a wall of text anymore.

Here's token-by-token streaming in https://t.co/PbpArqfiGd Core, start to finish. 🧵

[Sulay Joshi @sulayjoshi]

await foreach (var u in client.CompleteStreamingAsync(msg, ct))
{
await stream.WriteAsync(Encoding.UTF8.GetBytes(u.Text), ct);
await stream.FlushAsync(ct);
}

Cancellation + explicit flush matter more than the loop itself. #DotNetAI

[Sulay Joshi @sulayjoshi]

@dotnet From .NET 11 JIT optimizations to automated zone placement for VM Scale Sets, this episode hits every layer of the cloud-native stack. Bringing database SDKs into the multicloud era while securing the baseline is exactly what enterprise teams need.

[Sulay Joshi @sulayjoshi]

@dotnetfdn @realDotNetDave An absolute masterclass in open-source performance tuning! Shaving execution time down to get gains up to 87.52× across hundreds of methods is massive for enterprise .NET 10 workloads.

[Sulay Joshi @sulayjoshi]

@AzureDevOps The real skill ceiling in modern development isn't letting an AI agent spit out boilerplate code; it's active steering, prompt refinement, and strict architectural oversight. It’s an amplification tool, not an autopilot.

[Sulay Joshi @sulayjoshi]

var chat = new AgentGroupChat(researcher, writer);

Researcher gathers facts. Writer turns them into prose. Neither needs to be good at both.

When quality drops, you know exactly where to look.

#DotNetAI #CSharp

[Sulay Joshi @sulayjoshi]

Multi-agent systems aren't hype — they're just the Single Responsibility Principle applied to prompts. 🧵

[Sulay Joshi @sulayjoshi]

One agent that does everything = one giant prompt that fails silently, and you can't tell which part broke.

Here's a 2-agent .NET setup that actually works.

#SemanticKernel

[Sulay Joshi @sulayjoshi]

@dotnet Bringing native agentic UI components to the ecosystem is a massive win. Being able to build real-time streaming chat interfaces using familiar C# and Razor syntax instead of rewriting the front-end orchestration loop completely redefines developer velocity.

[Sulay Joshi @sulayjoshi]

@VisualStudio A classic Raymond Chen masterclass in engineering realities. It’s so easy to hunt for a quick optimization by swapping a class for a struct, only to get completely sabotaged by underlying copy-by-value semantics. A must-read for UI devs!

[Sulay Joshi @sulayjoshi]

@dotnetfdn Spot on. An agent is not just an LLM with tools bolted on. Deconstructing it into a controlled harness, reasoning engine, and strict directives is how we move from fragile demos to predictable, production-grade systems.

[Sulay Joshi @sulayjoshi]

No magic. The model recognizes it needs data it doesn't have, calls your function, reads the result, answers naturally.

Agentic behavior = a design pattern, not mystery.

#CSharp #AIEngineering

[Sulay Joshi @sulayjoshi]

This is how ChatGPT plugins actually work — and you can build the same thing in C# today.

Function calling, explained. 🧵

[Sulay Joshi @sulayjoshi]

[KernelFunction, Description("Gets weather for a city")] public string GetWeather(string city) => ...;

kernel.Plugins.AddFromType<WeatherPlugin>();

The model decides WHEN to call YOUR code.

#SemanticKernel #DotNetAI