cs | backend & systems dev | oss
India
Joined May 2021
- Tweets150
- Following180
- Followers100
- Likes937
Saksham retweeted
Meet Panoptik for the @OpenAIDevs WebMCP Challenge!
A browser-native video studio where humans & AI agents co-edit via 30+ WebMCP tools.
Live: panoptik-studio.vercel.app/e…
Continued below:
I kept running into the same issue while building AI agents: How do you safely execute untrusted AI-generated code on a server without losing control?
So I built Bastion: A sandboxed execution layer for AI agents to run untrusted code and manage files.
github.com/sakkshm/bastion
Core design:
- Each session = Docker-isolated runtime
- Persistent state
- Policy + resource controls (CPU, memory, network, filesystem, time)
- Concurrent execution inside sessions
- WebSocket terminal for bidirectional I/O
- Structured logs for observability/debugging
Bastion is not just a container runner.
It’s a stateful execution layer for AI-native systems with:
- isolation
- controllability
- persistence
- observability
Learn more about the project by reading the blog: sakkshm.me/blog/bastion-blog
Everyone’s building AI agents, but few talk about where they actually run.
The bottleneck isn’t the model, it’s execution. Once code is generated, there’s no reliable way to run it.
So I’m building Bastion: a runtime for executing agent code in isolated, persistent environments
🤖 Made with AI
Most setups are still:
LLM → generate code → run it somewhere → hope it works
There’s no clear execution model. No guarantees around isolation, state, or failure modes. It works for demos, not for real systems.
Got accepted into the Blok Capital Builder Cohort #1.
Excited to explore protocol design and web3 infra with an insane group of builders.
@blok_cap @0xSheetal @0xjayeshyadav @vansh_sahay
Saksham retweeted
Introducing SunoHQ, a platform helping small businesses launch AI voice agents in minutes.
Automate customer support on Telegram today, with WhatsApp & Calls coming soon.
No coding required, Regional languages and agents stay active 24/7.
@SarvamAI #TheMicIsYours @sakkshm
The entire pipeline is open source:
Python + DSP + hashing + indexed database lookups.
If you’re interested in how systems like Shazam actually work under the hood, the code is here: github.com/sakkshm/spectra
I’ve been trying to understand how Shazam-style song recognition actually works.
What surprised me is that it’s not ML-heavy at all. It’s mostly signal processing and a clever way to hash audio so you never have to compare raw waveforms.
Matching is basically an alignment problem. Many hashes will match many songs, but only the correct song produces a consistent offset between where the hashes appear in the sample and in the original track. That consistency is enough to identify a song from a few seconds.