ProjectsWork with meWritingAboutSay hi →
ON THIS PAGE01/05
  1. 01Why I built this
  2. 02The character
  3. 03What it does today
  4. 04What I'd expand
  5. 05Takeaway
ON THIS PAGE01/05
  1. 01Why I built this
  2. 02The character
  3. 03What it does today
  4. 04What I'd expand
  5. 05Takeaway
All writing
Wednesday — a Python voice assistant with a cold personality
Case study

Wednesday — a Python voice assistant with a cold personality

A personal voice assistant. JARVIS + FRIDAY as the tech inspiration, Wednesday Addams as the personality. Runs on Python and LLM APIs.

eBy eloi·November 18, 2025·2 min read
SHARE
Client
My own product
Role
Design · Full-stack · Python · Voice UX
Year
2025
Stack
Python · LLM APIs · TTS · STT · System prompts

Why I built this

I've watched Iron Man too many times. JARVIS and later FRIDAY were the AI assistants I always wanted — always-on, voice-first, dry, actually useful. Modern LLMs plus decent TTS/STT finally made that pattern buildable for one person.

Off-the-shelf assistants like Siri or Alexa are competent but generic. I wanted something with a distinct personality — not the eager helper, not the corporate-friendly voice. Something that felt like it had a mind of its own.

The character

Wednesday Addams. Cold. Dry. Direct. Doesn't sugarcoat, doesn't apologize when the answer is “you're wrong.” Occasionally deadpan-funny in a way that makes the interaction memorable.

It's enforced through the system prompt — no hedging, no unnecessary pleasantries, no “as an AI language model” preamble. If the answer is short, it's short. If the answer is that I made a mistake, that's what she says. And in voice, you actually hear that tone.

What it does today

Voice loop over LLM APIs. I speak, she listens, she answers back. Python for the loop, STT for capturing what I said, TTS for her voice.

Voice-first changes the constraints:

  • Responses have to sound right when spoken, not just read right
  • Long answers get truncated or summarized — nobody wants a monologue
  • Filler words like “actually” and “essentially” pile up quickly in TTS, so they get stripped from the prompt

What I'd expand

  • Long-term memory across sessions so she actually gets to know me
  • Tool use (calendar, email, code exec) — the jump from “answers” to “acts”
  • Wake word for hands-free trigger
  • Voice tuning — the character is 90% prompt, 10% voice, and the voice part still has room

Live-in-progress build, not a finished product.

Takeaway

The gap between generic AI and an AI with a distinct personality is smaller than I expected. Most of it is prompt engineering and voice-response design — rhythm, brevity, tone. Not model quality.

That's a good sign for personal tools. The intelligence is a commodity now. The character isn't.

NEWER POST

PDF Toolkit — a fully client-side toolkit that never uploads your files

Case study
OLDER POST

Conversion Landing Page — one clean page fixed visibility, trust, and bookings in a week

Case study
KEEP READING

You might also like

  • Case studyFeb 14, 2026

    Fury AI — a chat companion, named for my first cat

    A local LLM chat interface. Named after Fury, my first cat, who passed last year. I built it because I couldn't stop missing him, and typing to a screen felt closer than nothing.

  • Case studyMay 18, 2026

    DeepSite — website audits that give you the fix, not just the diagnosis

    Existing audit tools hand you a Lighthouse score and 47 tabs of docs. DeepSite scans your site once and returns a plain-English report you can actually act on.

designed, built & shipped
by one person.
Site
ProjectsWritingWork with meProcessAbout
Where
Philippines
Working worldwide
GMT+8 · async-friendly
Talk
eloi · made with care · © 2026
Privacy·Terms·Sub-processors
Available for new projects