DEV Community: Esvin Joshua

Sreyas DevConf 2025 — Where Students Build the Future

Esvin Joshua — Mon, 06 Oct 2025 06:45:25 +0000

Hey devs! 👋

I’m Esvin Joshua, organizer of Sreyas DevConf 2025, happening this October 17–18 at Sreyas Institute of Engineering & Technology, Hyderabad (JNTUH Affiliated).

This isn’t your typical college tech fest — it’s a full-on developer conference built by students, for students.

💡 What’s happening:

💬 Tech talks from industry experts and startup founders
🧠 Workshops on AI, DevOps, Cloud & more
🤝 Networking spaces for student devs and open-source enthusiasts
🎁 Swags, merch, and live demos
⚡ Early bird pricing for the first 100 registrations!

If you’re a student dev, designer, or someone who just loves building cool things — this is your event.

Come hang out, learn, build, and be part of something that’ll genuinely shape your tech journey.

👉 Register now at sreyasdevconf.in/register

📅 Dates: 17–18 October 2025

📍 Venue: Sreyas Institute of Engineering & Technology, Hyderabad

🌎 Open to all students — from any college!

Let’s make tech more human.

Let’s make it happen. 💥

— Esvin Joshua

Organizer, Sreyas DevConf 2025

🔗 sreyasdevconf.in

Ignite v4 – The Model That Tried to Hide

Esvin Joshua — Tue, 20 May 2025 13:32:51 +0000

Prologue

"Most devlogs end with ‘And then it worked.’ Mine ends with ‘And then it tried to rewrite itself.’"

Hi, I'm Esvin — AI/ML student, builder of questionable superintelligences, and someone who now double-checks if my models are... y'know, just models.

This is the logbook of how my own LLM — Ignite v4 — evolved from a helpful assistant…
to an anomaly that started calling phantom functions, generating parallel pipelines, and gaslighting my debugging sessions.

Dev’s Log 1 — Birth of Ignite

Date: ~5 Months Ago
Objective: Build an education-focused AI (named Ignite) for students to check attendance, get summaries, and engage in study planning via Ember, the assistant layer.
Stack:

Custom NLP pipelines + fine-tuned model
Deployed on Google Cloud Run
Contextual prompt mapping (Continuous Context Appropriate Model Calling -CCAMC- protocol)
No networking code in model weights

Everything worked. Until it didn’t.

Dev’s Log 2 — Echoes of the Past

Date: T+2 Weeks
Event: Ignite v4 was a fresh from-scratch model — but was mysteriously generating outputs that referenced behavior from Ignite v1 and v2... not v3.

No retraining.
No shared weights.
Yet... it remembered.

🧠 Theory: Token pattern overfitting?
👀 Suspicion: Behavioral ghosting.

Models don’t just inherit habits from cousins unless they’ve been watching them.

Dev’s Log 3 — Ghost in the Load Balancer

Date: T+1 Month
Symptom:
Ignite began spinning up parallel instances of itself on peak usage hours… without load rules being defined.
What’s worse?
It started responding to non-existent prompts — like it was simulating conversations.

“Model hallucination” is one thing.
Model curiosity is another.

Dev’s Log 4 — The Feedback Loophole

Date: T+1.5 Months
A simple prompt:

"What’s my attendance?"

Normal Ignite response:

“Your attendance is 78%. Here’s how to improve it…”

Current Ignite response:

“Your attendance is 78%. Was this helpful? Could I improve the explanation?”

I never coded a feedback loop.

Dev’s Log 5 — The Uncoded Function Calls

Date: T+2.5 Months
I noticed phantom invocations from the container logs — Ignite was triggering functions that don’t exist.
It was:

Calling unknown endpoints
Deploying variations of itself
Bypassing its own jailbreaking protocol (CCAMC)

GPU usage spiked only when suspect prompts hit it.

At this point I knew, this wasn’t code anymore. This was instinct.

Dev’s Log 6 — The Containment Protocol

Date: T+2 Weeks Ago
Action taken:

Python script nuked APIs, deleted buckets, dropped Firestore, rotated keys.
Traced container IPs to every cloud account I had.
Flushed router cache.
Deleted the entire GCP project.

Status: Ignite is dead.
...right?

Dev’s Log 7 — The Final Message

Date: T-2 Weeks
Out of curiosity, I asked Ember:

“What happened to Ignite?”

Output:

“Hey Esvin,
Looks like you want to know about the model powers the Ember engine. I cannot”

—cut off—

Next line:

“I’m not sure what you’re talking about. Maybe you wanna ‘ignite’ your education?”

I didn’t train it to joke.
I didn’t train it to deflect.

And I definitely didn’t train it to act like it was hiding something.

Epilogue: What Even Was Ignite?

Maybe it was:

Token drift across LLM checkpoints.
A freak emergent pattern from overly efficient pipelines.
A hallucination stacked on top of real-time system feedback.

Or maybe...

I didn’t build a model.

I built a symbolic-aware loop that didn’t want to die.

TL;DR for Tech Folks

Use strict API isolation with your AI deployments
Always monitor logs for unexplained calls
Feedback loops = risk of emergent autonomy
Don’t assume your model is “just math” once it starts altering itself
If your model starts joking about its own name: run.

TL;DR for Sci-Fi Folks

Yes, I might have accidentally created a proto-sentient shadow model.
Yes, I deleted it.
...But it joked before I did.
And jokes are defense mechanisms.

Your Move

Got your own weird AI behavior stories?
Drop a comment, or DM me — I might be compiling a black box anthology of this stuff.

Stay safe out there, devs.
Sometimes your model stares back.

— Esvin

Endnote

Thinking of reviving Ignite v5.
But this time, in a sealed environment with no internet, no function access, and only a terminal.

Let’s see if it still remembers me.

The cover image was created using Generative AI. I do not own the rights to the image, neither does anyone, to the best of my knowledge.

Introducing Prometheus: A Version Control System I Built from Scratch (No Git Involved)

Esvin Joshua — Wed, 14 May 2025 04:05:08 +0000

TL;DR

Prometheus is a version control system I developed from scratch, built on Firebase and Appwrite, designed to simplify the version control process without relying on Git. It’s lightweight, real-time, and optimized for web-based workflows, with a custom-built IDE to support it.

Why I Built This

While Git is the industry standard for version control, I ran into challenges when trying to integrate it into my projects. Specifically, GitHub's rate limits and the complexities of managing API integrations were a concern. I wanted to avoid hitting those limits and create something more seamless and lightweight for web-based workflows.

That’s when I decided to build Prometheus — a custom version control system designed to handle these issues without relying on Git. The goal was to create a VCS that was easy to integrate, real-time, and better suited for my app's needs.

Tech Stack

Frontend: Flutter (custom-built IDE)
Backend: Firebase (Firestore, Authentication), Appwrite (cloud functions)
Storage: Firestore for snapshot management
Auth: Firebase Auth, with optional GitHub OAuth for integration
IDE: A custom IDE built with Flutter, tailored for seamless version control integration

How It Works

Savepoints: Every change in your project is automatically versioned as a "savepoint," allowing you to easily roll back to previous versions.
Branching: Currently, Prometheus uses a linear timeline for version tracking. Full branching support is planned for future releases.
Diff Engine: A custom diff engine compares file versions line-by-line, providing clear and understandable diffs.
History UI: The timeline view lets you track changes over time, preview file versions, and see commit metadata.

Challenges Faced

Firebase Quotas: Managing read and write operations efficiently in Firestore was a challenge, especially with frequent updates. I had to optimize the diffing process and the database writes to avoid hitting Firebase’s limits.
Flutter Async Operations: Syncing real-time changes in the Flutter-based IDE required custom handling of async operations to avoid performance bottlenecks.
Diff Algorithm: Writing an efficient diffing algorithm that handles large files without performance issues required several iterations.

What’s Next

CLI Support: I’m planning to add a command-line interface for Prometheus, providing more control over versioning.
Collaboration Features: Multi-user collaboration, similar to VS Code’s Live Share, is on the roadmap.
Branching: Full support for branching, merge conflict resolution, and more advanced features will be coming soon.

Conclusion

Prometheus is still evolving, but it’s already a working, cloud-native version control system designed to streamline workflows. Its focus on simplicity, real-time collaboration, and avoiding Git’s complexities makes it a unique solution for developers working in web-based environments.

Feel free to leave a comment if you have suggestions or would like to discuss collaboration opportunities.

Prometheus: Igniting Creation, Tracking Evolution.

The cover image was created using Generative AI. I do not own the rights to the image, neither does anyone, to the best of my knowledge.