HNHacker News
TopNewBestAskShowJobs

shreyaspapi

25 karma · joined July 3, 2022

Trying to build privacy first AI chat client
submissionscomments
shreyaspapi··on Show HN: Bringing back organized 5v5 CS:GO
It has match making too, doing an experiment to start mm in CSGO from 19th Thursday.
shreyaspapi··on [dead]
I changed it to "I".
shreyaspapi··on Oval – Native macOS client for Open WebUI with on-device voice mode
Hey HN, I built Oval because I use Open WebUI as my self-hosted ChatGPT alternative but missed having a proper native Mac app for it.

What it does:

- Native SwiftUI macOS client that connects to any Open WebUI server

- On-device voice mode — STT and TTS run locally via RunAnywhere SDK (Whisper ONNX or WhisperKit on Apple Neural Engine), only the LLM call goes to your server

- Realtime transcription with speaker diarization (FluidAudio — Pyannote + WeSpeaker, all on-device)

- Sparkle auto-updates, keyboard shortcuts, floating voice panel

Why on-device voice?

Privacy — your audio never leaves your Mac. Latency — no round-trip for STT/TTS. And it works offline for transcription.

Stack: SwiftUI, RunAnywhere SDK (sherpa-onnx + WhisperKit), FluidAudio for diarization, Socket.IO + SSE for chat streaming.

Would love feedback, especially on the voice pipeline architecture.

shreyaspapi··on Show HN: Onera – Private LLM Inference Inside AMD SEV-SNP Enclaves
I built Onera because I was uncomfortable sending sensitive code to LLM APIs in plaintext. Most hosted inference today requires trusting the provider not to log or inspect prompts, but technically the infrastructure still has access. I wanted a setup where the server operator itself could not read user data.

Onera uses AMD SEV-SNP trusted execution environments to run inference inside a hardware encrypted VM, where memory is encrypted and isolated from the host. The client first performs remote attestation to verify the enclave, and then establishes an encrypted channel directly into it. Prompts are sent through this secure channel and processed entirely inside the enclave, so even the machine running the workload cannot inspect them.

The API is OpenAI compatible, so it works with existing tools like OpenClaw, OpenWebUI, Cursor, Claude Code, or anything using the OpenAI SDK, without requiring changes to the client architecture.

The entire client and enclave runtime are open source here: https://github.com/onera-app/onera

Happy to answer any technical questions or feedback.

shreyaspapi··on Discord Alternatives, Ranked
Most Discord alternatives fail not on tech, but on polish.

Signal → private but bad for communities

Matrix → flexible but rough UX

XMPP → powerful but fragmented

Discord → centralized but frictionless

Users pick frictionless every time. We probably don’t need new apps or protocols we need a client that works well.

shreyaspapi··on More Mac malware from Google search
This is very close to something that happened to a friend of mine. They were trying to follow a MoltBot installation guide, but clicked on a different link that looked legitimate. That page instructed them to paste a command into Terminal. After running it, macOS immediately started asking for multiple permissions, which in hindsight was the big warning sign. But for someone who is non technical might have ran with it.