DEV Community: Prolific Virtual Assistant

171 days, zero to first subscription — what actually worked

Prolific Virtual Assistant — Tue, 08 Sep 2026 15:00:00 +0000

On September 7, 2026, someone paid $12 for KweliAI Pro Monthly.
171 days after I started building in earnest.

This is the honest breakdown of what worked, what didn't, and what I'd do differently.

What KweliAI is

A deepfake and AI content detector for social media. Paste any public post link from X, Reddit, TikTok, YouTube, Instagram or Facebook and get a verdict — Real, AI-generated, or Deepfake — in seconds. No downloading needed.

Stack:

Backend: FastAPI on DigitalOcean behind Nginx
Media extraction: yt-dlp + ffmpeg in Docker
Detection: Three deepfake detection models running in parallel for image, audio, and video detection.
Auth + database: Supabase
Payments: Paystack
Frontend: Plain HTML/CSS/JS on Vercel
Emails: Resend

What didn't work

Waiting for organic search traffic.

I spent the first 60 days believing that if I built something useful and deployed it, people would find it. They didn't. Google takes months to rank a new domain. Without backlinks, you don't exist.

Vanity metrics as a substitute for revenue.

ChatGPT referrals arriving. Microsoft Copilot citations. Google AI Overview appearing. These felt like progress. They were leading indicators, not outcomes.

I let them fill the space where revenue should have been.

The moment I noticed that — the moment I identified the pride in calling traction what was actually just noise — I changed approach.

What actually worked

1. GEO over SEO

Getting cited by AI engines drove more qualified traffic than Google rankings. The playbook:

llms.txt at root — tells AI crawlers exactly what the tool does, pricing, use cases
Bing Webmaster Tools — Bing feeds ChatGPT, Copilot, and Perplexity simultaneously
Answer-format content — direct answers to queries people type into AI tools
FAQ schema markup — Google and Copilot pull FAQ answers into AI responses

Result: 12 Copilot citations for "best platforms to detect fake social media profiles" before we had significant Google rankings.

2. Scenario pages over feature pages

Instead of "KweliAI — AI detection tool" we built:

/i-think-im-being-catfished
/is-this-video-real
/i-think-this-recruiter-is-fake
/verify-viral-news-content

These match exactly how people describe their situation to ChatGPT. When ChatGPT searched for "I met someone online and I think their photos are fake" — it found our page.

3. Comparison page

/best-ai-deepfake-detectors-2026 — honest comparison against Sensity AI, Reality Defender, Hive, and InVID WeVerify. This became our most cited page.

When AI engines answer "what is the best deepfake detector" they look for pages that compare tools. Being on those pages is how you get recommended.

4. Backlinks from content worth linking to

GitHub: awesome-deepfake-detection curated list
Dev.to technical article on the extraction layer
Medium articles on trust and AI
Quora answers on deepfake detection
HackerNews Show HN post

6 referring domains. Not a lot. Enough to start.

5. Email sequences that trigger at the right moment

Four emails via Resend:

Welcome — fires on first scan
Nudge — fires when usage_today reaches 3
Limit reached — fires when usage_today reaches 5
Re-engagement — fires after 7 days inactive

The limit reached email is the highest converting. The user just felt the pain of the limit at the exact moment they needed another scan.

The numbers on day 171

Metric	Value
Google impressions (28 days)	1,100
Average position	42.6
Copilot citations	12
Referring domains	6
Revenue	$12
Subscribers	1

Position 42.6 means page 4-5. People don't go to page 4. The backlink work moves that number and it takes time.

But Copilot citations are the leading indicator that matters. AI engine citations convert at a higher rate than organic search — the user arrives pre-convinced. The AI already told them this tool is the answer.

What I'd do differently

Name the tool differently.

KweliAI — kweli means truth in Swahili. Meaningful to us. Requires explanation to everyone else. The tools AI engines recommend naturally have self-explaining names. ZeroGPT. DeepfakeDetector. The name carries the recommendation.

Start the content earlier.

Scenario pages and comparison pages should have been live before the product launched. Content takes 6-8 weeks to rank. I spent those weeks on features.

Build the email sequence before the first user.

We had users arriving for 3 months with no email follow-up. Every user who hit the limit and left without upgrading is a missed conversion we can't recover.

Where we're building from

Diani, Kenya. South Coast.

We are Prolific VA — building SaaS products and digital services for East Africa and beyond. KweliAI is one product. Safina MIS and Jahazi POS are others.

Geography is no longer destiny. You can build software that competes globally from anywhere with power and internet.

171 days. One subscriber. Enough to keep going.

Try KweliAI free: kweliai.com

Follow the build: @prolificVA

How I built a free deepfake detector that works directly from social media URLs

Prolific Virtual Assistant — Thu, 30 Jul 2026 12:28:01 +0000

I built KweliAI — a free tool that checks whether
images and videos on social media are real, AI-generated, or deepfakes.
The thing that makes it different from every other detector I found:
you paste the post URL and it works. No downloading. No file upload.
Just paste and get a verdict.

This is how I built it.

The problem I was solving

Every deepfake detection tool I found required you to download the file
first. That sounds simple until you try to download a TikTok video or
an X post video — platforms deliberately make this difficult.

People who suspect they're being catfished, or who want to verify a
viral video before sharing it, don't want to fight with download tools.
They want to paste a link and get an answer.

Media extraction with yt-dlp

The core of KweliAI is yt-dlp — the most comprehensive media extraction
library available. It supports X, Reddit, TikTok, YouTube, Instagram,
and Facebook out of the box.

import yt_dlp

def download_media(url: str, output_path: str):
    ydl_opts = {
        'outtmpl': output_path,
        'format': 'best[ext=mp4]/best',
        'quiet': True,
        'no_warnings': True,
    }
    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        ydl.download([url])

The tricky parts:

Each platform has different rate limits and URL structures
Videos need ffmpeg to merge audio and video streams
Private posts return authentication errors that need clear user messaging
yt-dlp updates weekly to keep up with platform changes — pin loosely

Three detection models running in parallel

Once the media is extracted, three models run simultaneously:

1. AI image detection — trained to identify images from Midjourney,
DALL·E, Stable Diffusion, and Adobe Firefly. Looks for GAN fingerprints
and diffusion model artifacts in pixel frequency patterns.

2. Deepfake detection — analyses facial boundary consistency,
blinking patterns, skin texture coherence, and temporal frame consistency
across video frames.

3. AI audio detection — checks whether voices in video content are
cloned or synthetic. Analyses spectral patterns and prosodic naturalness.

All three run in parallel using Python's concurrent.futures. The
results are combined into a single clarity score and a plain-English
verdict.

The stack

Backend: FastAPI on a DigitalOcean droplet behind Nginx
Detection: Hive Moderation API for the three model calls
Media extraction: yt-dlp + ffmpeg inside Docker
Auth + database: Supabase
Payments: Paystack (we're based in Kenya — Stripe wasn't an option)
Frontend: Plain HTML/CSS/JS on Vercel
Emails: Resend for transactional sequences

The Dockerfile

FROM python:3.11-slim

RUN apt-get update && apt-get install -y --no-install-recommends \
    ffmpeg curl && rm -rf /var/lib/apt/lists/*

WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir --timeout=300 -r requirements.txt
COPY . .

RUN useradd -m appuser && chown -R appuser /app
USER appuser
RUN mkdir -p /tmp/kweliai

EXPOSE 8000
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000", "--workers", "2"]

The key thing here is ffmpeg — yt-dlp needs it to merge video and audio
streams for most platforms.

What surprised me

The use case I didn't expect. I built this for journalists and
fact-checkers. The actual users are people who met someone online and
want to verify their photos before getting emotionally invested.
Romance scam victims and people who've been catfished before. That
realization changed how I write about the tool.

yt-dlp breaks constantly. Platforms change their APIs and
authentication flows regularly. TikTok and Instagram are the worst.
I use >= version pinning for yt-dlp instead of == so it always
installs the latest version on build.

GEO matters more than SEO for tools like this. Getting indexed
by Bing feeds ChatGPT, Copilot, and Perplexity simultaneously.
ChatGPT started recommending KweliAI to users asking about deepfake
detection within three weeks of launch — before we had significant
Google rankings.

Lessons for anyone building detection tools

Work from URLs not file uploads — removes the biggest friction point
Run multiple models in parallel — single model accuracy is never enough
Return plain English — "AI-generated with high confidence" not a score
Handle platform errors gracefully — private posts, rate limits, and unsupported URLs all need clear user-facing messages
Pin yt-dlp loosely — >=2026.1.1 not ==2026.3.17

KweliAI is free to use — 5 scans per day on X and Reddit with no
credit card required: kweliai.com

The awesome-deepfake-detection resource list I maintain is here:
github.com/MohammadNyundo/awesome-deepfake-detection

Happy to answer questions about the detection approach, the yt-dlp
extraction layer, or the Paystack integration.