A dark dashboard showing a voice assistant interface with memory timeline

03 / 06 — AI Assistant

A personal assistant with memory, voice, and brakes.

Client
Personal project
Year
2026
Role
Architecture & engineering
Stack
Python · Local STT/TTS · Vector Memory · WebSockets

Overview

Helix listens, remembers, and answers in the language it was spoken to. It is designed to run continuously on modest hardware at effectively zero marginal cost.

Helix dashboard — voice waveform, memory timeline, and action log
Fig. 1The live dashboard. See what it's doing.

01 — The challenge

An assistant with long-term memory and the ability to act on your behalf is a different risk category from a chatbot. The hard part is not the intelligence — it is giving it enough capability to be useful while keeping it constrained enough to trust.

02 — The solution

Multilingual voice with automatic language detection, semantic memory that embeds conversations and extracts facts without manual filing, and a live dashboard over WebSockets so you can see what it is doing rather than guessing. Around that sits a layered safety model — execution sandbox, append-only audit log, rate limiting, and a kill switch.

4-core

Runs on a cheap ARM server

0

Cloud dependencies

24/7

Always listening

Helix memory view showing extracted facts and conversation embeddings
Fig. 2Semantic memory — facts extracted automatically.

Credits

  • Architecture & buildRahul Kumar Singh

Next project

Pilot BrowserA browser driven over raw CDP — for pages that fight back.
A browser automation interface showing CDP commands and a live page preview