Konstantin Tarkus

Building the rails that AI workflows run on. Principal Engineer architecting the foundation for safe & scalable intelligence. ☕️ // Code #DevOps #Startups #AI

[Konstantin Tarkus @koistya]

This data backs up exactly what we do in practice.

Instead of a massive AI-generated repo overview, our React Starter Kit uses an https://t.co/BzkNtWK1wQ that is strictly under 50 lines of code. Check it out here: https://t.co/kdmTb8NcZr #ReactJS #LLMs

[Konstantin Tarkus @koistya]

Are https://t.co/BzkNtWK1wQ files actually helping your AI agents? 🤔

A new evaluation shows auto-generating context files actually decreases task success by 3% and increases costs by 20%+!

The takeaway? Write them yourself and keep them strictly minimal. 🧵👇

[Verist @verist_ai]

🚀 Verist v0.0.1

Verist is a determenistic workflow kernel for AI systems – built for replay, recompute, and diff.

→ No agents.
→ No hidden state.
→ No silent overwrites.

If AI decisions need to be explainable months later, this is for you.

https://t.co/2OghPKB1wP

[Konstantin Tarkus @koistya]

Most APIs still ship their own SDKs, auth flows, and quirks.

But now many expose an MCP server.

So… why not just generate a typed client from that?

Same mental model. Any MCP API.

[Konstantin Tarkus @koistya]

MCP + OAuth + CLI apps = lots of sharp edges.

oauth-callback v2.2.0 smooths them out — redirect capture, token exchange, persistence.

https://t.co/bXxSGyhsY8

[Konstantin Tarkus @koistya]

@Achu_Ashwath_ @ThePracticalDev I personally use it daily to get targeted feedback on specific parts of the codebase from ChatGPT, Gemini, and Grok. Works like a charm! :)

[DEV Community @ThePracticalDev]

Struggling to get ChatGPT to understand your full codebase? This dev shares their bundling strategy to provide better context for AI assistance. 📦🧠

{ author: @koistya }

https://t.co/iTofgJRv8v

[Konstantin Tarkus @koistya]

Cloudflare vs. Vercel just got real. 🥊

☁️ Cloudflare = Global Infra 🚀 Astro = Zero JS / Peak DX

https://t.co/MDO9G5MwdX

[Konstantin Tarkus @koistya]

Cowork is a strong signal of where agentic workflows are heading.

I've been exploring similar ideas with @sidebandtech — focusing on E2EE-first, trust-bounded workflows where files don’t have to leave the user’s control.

#Claude #ClaudeCode #AI #AgenticAI

[Claude @claudeai]

Introducing Cowork: Claude Code for the rest of your work.

Cowork lets you complete non-technical tasks much like how developers use Claude Code.

[Konstantin Tarkus @koistya]

Released @sideband/secure-relay v0.1.0 – E2EE for browser↔daemon communication via untrusted relays.

• Ed25519 identity signatures
• X25519 ephemeral key exchange
• ChaCha20-Poly1305 encryption
• TOFU identity pinning
• Replay protection

https://t.co/USb1QzaqGg

[Konstantin Tarkus @koistya]

https://t.co/QWPIDNZwo2

[Konstantin Tarkus @koistya]

Just launched AI Assistant for React Starter Kit 🤖⚛️

It understands the actual codebase, architecture, and patterns, not generic React advice.

Built to help you ship faster with fewer “where does this go?” moments.

#React #TypeScript #DevTools #AI

[Konstantin Tarkus @koistya]

@r0ktech Using Astro for marketing pages + React/Vite/TanStack for the web apps (main app, tenant console, etc.)

Check out React Starter Kit on GitHub as an example.

[Konstantin Tarkus @koistya]

Best talk I've seen on the practical reality of coding with LLMs (and why "clean code" definitions are changing)
https://t.co/PwQt4lBu9v

[Konstantin Tarkus @koistya]

@DrizzleORM Great work! Happy Drizzle ORM user here.

[Drizzle ORM @DrizzleORM]

We've updated our benchmarks 🚀

Drizzle v1.0.0-beta2 vs Prisma 7.1

Node 24.6:
- Drizzle handled 2.13m requests, peaked at 9.4k reqs/s at ~100ms p95 latency
- Prisma handled 668k requests, peaked at 2.3k reqs/s at ~150-200ms p95 latency before throttling

Bun 1.3.4:
- Drizzle handled 2.33m requests, peaked at 11.6k reqs/s at 25-30ms p95 latency
- Prisma handled 678k requests, peaked at 2.3k reqs/s
at ~150-200ms p95 latency before throttling

[Konstantin Tarkus @koistya]

@bunjavascript Anyone willing to help integrating fake timers in this project? https://t.co/O8Du010hEu

[Konstantin Tarkus @koistya]

No DDoS, no hack. Cloudflare outage = one small ClickHouse ACL change → duplicate features → config file too big → Bot Management module hits preallocated memory limit → entire frontline proxy crashes repeatedly. Bravo to #Cloudflare team for transparency 👏