Systems & AI infra engineer. I write about Postgres, Kafka, Rust, and LLM systems - how real systems fail, scale, and get fixed Building @coldbrewai
India
Joined March 2015
- Tweets796
- Following611
- Followers85
- Likes620
Hey !
we have built @coldbrewai , Best AI app builder, which can help you create, manage and deploy apps. We also support creating agents for your day to day task.
Every app created in coldbrew can be connected with claude, chat gpt etc, to ease your day to day job.
coldbrew.ai
Jagdish Parihar retweeted
Introducing Rafa Alvarez: 50 years of cast iron, zero years of computers. 🍳☕
One sip later, his whole shop is online. 184 orders. 12 stores.
Made by hand. Sold everywhere.
Build with Coldbrew → coldbrew.ai
#Coldbrew #Ecommerce #SmallBusiness #MadeByHand #Anime
Built is-ai-down.oncoldbrew.ai from a single prompt on coldbrew.ai.
It shows live status for the major AI tools in one place, so you're not refreshing five status pages when your agent stops responding.
Prompt -> live site. That's the whole pitch.
Jagdish Parihar retweeted
Introducing Sabrina Coldwell, the pharmacist who never lets anyone slip through the cracks. 💊☕
412 patients. Every refill, every check-in, every journey, tracked.
Build with Coldbrew → coldbrew.ai
#Coldbrew #Healthcare #AIAgents #Pharmacy #animes
Jagdish Parihar retweeted
Introducing Duke Coldwell, Real Estate Mogul.
He sold 12 homes this week without sitting up.
Why work hard? Build smart.
Build with Coldbrew → coldbrew.ai
Jagdish Parihar retweeted
vibe coding is solved. vibe shipping isn't.
we've all been there:
built the app in a weekend ✨
demo slaps
someone says "can we use this for real?"
and now you need:
— auth that isn't a TODO
— a database the AI can't nuke
— permissions (your users can see each other's data rn)
— agents with actual guardrails
— workflows that survive a crash at 3am
that last 10% is 90% of the work. it's why your beautiful prototype is still a prototype.
so we built Coldbrew differently: the AI never writes the plumbing. sign-in, Postgres + managed migrations, role-based access, agent harness — already built, battle-tested, can't be improvised into oblivion by a model having a bad day.
your prompt only shapes the part that's actually yours — the business logic. and it stays promptable forever. change it by saying it.
built an online store with search, cart, payments + an AI support agent in one prompt. it's been running for weeks. real orders, real logs, agents scoped like employees.
your first app is free → coldbrew.ai
vibe code it. then actually ship it.
Jagdish Parihar retweeted
You’ve spent years learning how your industry works.
You know the exceptions, the bottlenecks, and what everyone else misses.
That expertise is the starting point for better software.
Tell Coldbrew how the work should happen. We’ll turn it into an app your team can log into and use.
Bring your domain expertise. Build with Coldbrew.
coldbrew.ai
my brain has a mixture-of-experts architecture and today the gating network routed 100% of tokens to the expert going 'this is so cool' and 0% to the one going 'we are so back / so cooked' — anyway I automated half my backend today thanks to fable and astra
Every new agent session starts with amnesia.
You spent an hour yesterday deciding not to use the queue-based approach. Today the agent proposes it again, confidently, with the same reasoning you already rejected.
The obvious fix is pasting a summary at the start. It doesn't work. A summary tells the agent what happened. It doesn't tell it what was decided, or why the alternative lost.
What works is a file the agent appends decisions to and reads before it plans.
Not a transcript. Entries like:
Chose X over Y because Z. Revisit if Z changes.
Three lines each. It writes one when a decision gets made, reads the file before proposing anything.
A transcript ages badly and a decision doesn't. "We discussed queues" is noise in a week. "Rejected queues because we never needed the ordering guarantee" is still true.
Context is what the model needs to think. Memory is what it needs to stop re-litigating.
Every long agent session dies the same way.
Not a crash. The context fills up, something compacts it, and the agent quietly loses the receipts: which command it ran, what the raw output actually said.
What survives is the summary. "Tests pass, migration applied."
What's gone is the proof.
So it re-runs work it already finished. Or worse, it defends a claim it can no longer check, because the only evidence it has left is its own earlier sentence saying the thing was true.
Now the important outputs go to a file the agent can re-read. The window holds reasoning, not storage.
A context window is working memory. It was never meant to be the log.
The prompt that stopped my agents from bluffing.
Old prompt: "Add retry logic to the upload handler."
Agent: writes it, reports "Done, added retries with exponential backoff." Never ran it. The retry branch had a bug. I found it in review, twenty minutes later.
New prompt ends with a verification contract:
1. Run the code path you changed. Not the whole test suite, the actual path.
2. Paste the raw output. Not a summary of it.
3. If you can't run it (no creds, no sandbox, no fixture), say exactly what you couldn't run.
4. If output and expectation differ, fix and re-run before you report.
What changed: the agent stopped treating "I think it works" as done. It hits the bug itself, fixes it, and I get a diff plus a log instead of a claim.
Rule 3 does the heavy lifting. Without a sanctioned way to say "I couldn't verify this", the model fills that gap with confidence.
We use multiple AI models at work, so I built a small dashboard to track whether services like Anthropic, OpenAI, xAI, and Google are up or down.
I made it as a fun side project and would love to know whether anyone else finds it useful.
is-ai-down.oncoldbrew.ai/
Jagdish Parihar retweeted
What does a production-ready Coldbrew app give you out of the box?
🔒 Enterprise SSO
🤖 Native MCP support
🔑 Pre-configured Agent Auth
🐘 Postgres + managed migrations
⚡ Async hooks & workflows
⚙️ Built-in Admin Portal
Built for AI-native production from day one.
Open two psql sessions.
Session A: BEGIN; DELETE FROM users; — 1M rows gone
Session B: SELECT count(*) — still 1,000,000
No locks. No waiting. The rows are deleted AND not deleted.
I built the machinery that makes this work — in 100 lines of TypeScript:
blog.jatin510.dev/mvcc-expla…
I used to think DB connection pooling was deep infrastructure magic.
So I tried to actually understand it — and it turned out to be two arrays and a counter:
→ connections waiting for work
→ work waiting for connections
Opened pg-pool's source to check... it's literally the same two arrays (_idle, _pendingQueue).
Wrote up the full breakdown, from sockets up: blog.jatin510.dev/connection…
Building agent systems has changed how I think about software.
As engineers, we spend years designing APIs.
But agents don’t really care about APIs.
They care about the environment around them:
Can I execute code?
Can I run tests?
Can I inspect logs?
Can I search the codebase?
Can I recover from failures?
A good example is tools like Codex TUI.
The model isn’t useful just because it can generate code.
It’s useful because it’s operating inside an environment where it can read files, run commands, execute tests, observe results, and iterate on failures.
The most capable agents I’ve seen weren’t using the smartest models.
They were operating in the best environments.
Models score insanely well on benchmarks.
But the moment you put them into real systems, failures start appearing everywhere:
unclear requirements, unstable execution environments, poor tool contracts, weak testing frameworks, inconsistent state, retries, long-running workflows.
And all of that eventually reflects in the output.
One thing I’ve realized while working on agentic AI systems:
a raw LLM is just potential.
The real capability comes from the harness around it —
execution environments, memory, context propagation, tool orchestration, sandboxing, observability, and recovery from partial failures.
A good harness can make even smaller models feel incredibly powerful.
A weak harness can make the best models unreliable.
#llm #agentHarness #sandbox
One thing I’m realizing while building agentic systems:
generating code is the easy part.
The hard part is the harness around the model —
sandboxed execution, retries, context propagation, tool isolation, state versioning, observability, and recovering from partial failures.
Once agents start executing code and mutating state, you’re no longer building a chatbot.
You’re building a distributed system around non-deterministic intelligence.
The model generates intent.
The runtime decides whether that intent becomes reliable software.
My mental model for building AI systems has changed a lot over the last year.
I started with prompts and LLMs.
Then tools, workflows, and “master agents”.
Now it feels like the real shift is happening around skills + execution environments.
We’re moving from deterministic software →
to systems where the agent does more of the reasoning and execution.
Better models expand capability.
Better systems define boundaries.
That balance is going to matter a lot.
#AI #LLM #Agents
Get 1 free month of Bevel Pro 🎉
Use code YRHR5G when you sign up or enter it in Settings.
join.bevel.health/YRHR5G