Gowtham Sarveswaran

Cloud Arch

[Matthew Berman @MatthewBerman]

My full interview with Tibo (@thsottiaux)

0:45 Tibo's Lessons from Google DeepMind
4:22 Building OpenAI’s Relentless Culture
7:23 Astra & Next Gen Models
11:18 How Fast AI Changes Developer Workflows
14:27 ChatGPT & Codex Merging
20:25 OpenAI vs. Anthropic
23:37 Why OpenAI Keeps Resetting Limits
30:25 Recursive Self-Improvement
32:00 Dangers That Caused "The Pause"
34:13 Will Ultra Fast Become the Default?
43:20 Why Everyone Needs to Try AI

[Gowtham Sarveswaran @GothamSarves]

Source code, architecture, reproducible setup, and retained evidence:
https://t.co/e8VC3KAuqS

[Gowtham Sarveswaran @GothamSarves]

Built Nightwatch for #AllThingsAgenticHackathon.

It detects a failing Gemma model, uses Gemini 3.6 Flash + Google ADK to design and train one bounded repair, then lets deterministic evidence, not the agent, decide whether it ships.

Demo: https://t.co/NmkMW0ZHCj

[Gowtham Sarveswaran @GothamSarves]

Most scams do not start with malware. They start with pressure.

A delivery notice says your package is held. A bank alert says your account is locked. A recruiter promises easy money. A “family member” says they lost their phone and needs help right now.

That is the moment I wanted Jawbreaker to help with.

Jawbreaker is my submission to Hugging Face Build Small Hackathon 2026. It is a small-model scam defense app for the moment before someone clicks a suspicious link, replies to an impersonator, shares a code, or sends money.

The idea came from a real family problem, with private details removed: scam messages that look urgent, personal, and plausible enough that someone might act before asking for help.

Paste a suspicious text, email, or DM, and Jawbreaker turns it into a plain safety card:

🍬 what the risk is
🍬 who the sender is pretending to be
🍬 how the message is pressuring you
🍬 what they want
🍬 what could happen
🍬 the safest next step
🍬 a note you can copy to someone you trust

Under the hood:
🍬 MiniCPM5-1B + custom Jawbreaker LoRA adapter
🍬 Gradio Space on Hugging Face ZeroGPU
🍬 Modal A100 for training and eval
🍬 OpenAI Codex for implementation, eval design, UI iteration, and submission docs
🍬 Public model, dataset/evals, article, and repo
🍬 Final hard eval: 632 cases, 0 dangerous undercalls, 0 unsafe actions, 0 invalid JSON

Not claiming it catches every scam. Just trying to make the moment before someone acts a little safer.

Live app: https://t.co/8QN8oi4WRz
Writeup: https://t.co/hUHLkszpxM
Model: https://t.co/knKe9Eaxau
Dataset/evals: https://t.co/MiuVhUYF58

Would love feedback, especially on tricky scam messages it should handle better.

Built with @huggingface , @Gradio , @OpenBMB , @modal , and @OpenAI Codex.

Demo: https://t.co/br456WOX4A

#HuggingFace #Gradio #BuildSmallHackathon #SmallModels #AI #CyberSafety #OpenSource #Modal #MiniCPM