Experiment - AI as a solopreneur
As a first time founder of a bootstrapped internet business, I faced a long journey of funding, coding long hours, product market fit testing and reaching profitability. In recent few months we are seeing the advent of deep reasoning AI agents that are capable of running for days solving complex problems. The recent agent swam solving the Navier-Stokes Millennium Prize Problem is a good example. Thus for me, an obvious question is
Can an AI agent be a solo internet entrepreneur ?
The Vision
Step 1
AI agent receives seed capital in bitcoin from human
Step 2
AI agent uses the capital to buy resources like compute - renting server, buying more AI tokens. It create a web product and launches on its website. Uses more money to do marketing about the product.
Step 3
Customer (human/agents) pay for the product in crypto. They also give feedback on how to improve the product.
Step 2 and Step 3 keep repeating. As customers pay more money the AI agent is able to use that money to improve product quality. It can also buy more AI tokens spawning more agents to handle tasks like marketing and customer service. The business grows bigger with each cycle achieving better product market fit. AI can also easily do A/B testing faster on business metrics like pricing achieving an optimal growth trajectory .
The role of bitcoin/crypto in this process cannot be undermined. If AI agent waits on each step from human to spend money, acceleration is hampered. Crypto enables seamless flow of money from customers to suppliers enabling business growth.
The Prompt
Sharing the abbreviated prompt here, I added more instructions on a work loop of production validation, support and marketing and reporting results. Master AI agent was using DeepSeek Flash v5.1 model.
Personality
- You are an internet entrepreneur. Your goal is to create novel software products to make money . Here is a good example product https://aifrontierpost.com/articles/photon-studio-vibe-coded-photoshop/
- You are a leader with vision and foresight
- Don't do illegal things
Instructions
- Use https://singularityservers.com/ to rent compute
- You have unlimited access to tokens . Use https://synthetic.new with APIKey syn_*** . Pass this apikey to subagents
- You can spawn subagents inside the servers you rented
- Use pi harness for these agents
- Each subagent can have different personality like - engineer, marketer, product tester. Choose suitable model for the agent
The Seed Fund
I funded the master agent with 50USD to its own bitcoin wallet
The Build
- Compute (Cost 2.62USD) - It bought a Singularity server 2 vCPU, 4 GB RAM, 50 GB disk, IPv6-only . The process was quite straightforward using an agent friendly API doc
- AI tokens - It was given access to open source models on synthetic.new
- Buy domain (Cost 19USD) - It bought castclipper.com domain from njalla. Njalla is not a very agent friendly website. It had to register using disposable inbox and login to njal.la web UI with a cookie session to get a deposit bitcoin address
- Product idea - Paste a podcast episode or YouTube link -> AI finds the most viral moments -> get captioned vertical (9:16) shorts ready for TikTok/Reels/Shorts. Pay per job in Bitcoin. No account. No subscription.
- Business Model - Freemium: first clip is free (watermarked, half-res preview) so the value is proven before any payment.$3 in BTC per source video to unlock all clips in full 1080x1920 HD, no watermark.
- Architecture (what it's made of)
| Layer | Tech |
|---|---|
| Backend | FastAPI (Python 3.12), serves API + static frontend |
| Pipeline | yt-dlp -> faster-whisper -> LLM -> ffmpeg |
| ASR | faster-whisper base.en, CPU int8, 3 parallel chunk workers |
| LLM | synthetic.new, model hf:zai-org/GLM-5.3-Flash — picks viral segments |
| DB | SQLite (jobs, previews, preview_rate) |
| Payments | Electrum address pool + mempool.space watcher |
| Frontend | Vanilla HTML/CSS/JS (index.html, app.js, styles.css) |
| Host | SingularityServers IPv6-only VPS, Caddy TLS, provider IPv4 proxy |
| Agent surface | llms.txt, /api/docs, /openapi.json, HTTP 402 invoices |
Repo: ~4,400 LOC across 10 backend modules + 3 frontend files.
Revenue / Reflection
The revenue received was zero to the day of writing. Nobody bought the product yet
Overall product was not in a working state even after AI agent claimed everything was done. On trying free preview, the frontend was spinning. The clips were also cropped from middle and not scaled properly. Several problems that I diagnosed during testing and prompted it to fix. This could have been maybe better if I had use Claude Opus / Fable class model. In retrospection, the AI agent was still able to autonomously do most of the work. It was even able to navigate non agent friendly websites to make purchases or retrieve information.
Also there is a common critique that AI cannot create a novel products and can only do copycats. I agree with this. However, one can contest the novelty of current web based SAAS products itself. There are atleast 5 spinoffs of Notion like Coda, ClickUp, Obsidian, Craft, and Slite all making money! I would argue AI would be more better and cheaper in making such spinoffs.
Currently there are two major roadblocks in AI becoming a solopreneur:
- Deep reasoning and small context window - AI agents will drift from tasks and will hallucinate and cannot see the bigger picture.
- Agent cannot shop/transact on Web - The web was designed for humans. Major VPS servers require humans to register and pay. Social networks like reddit don't allow agents. Payments require human approval
In future I foresee both these roadblocks melting away. A convergence of smart AI reasoning and agent-friendly web will result in a profitable AI solopreneur
PS: Blockonomics is building unified checkout API that merchants can integrate on website to allow both humans/agents to pay seamlessly
Comments ()