The customer’s problem

Young e-commerce brands run on data spread across Shopify, GA4 and their ad platforms, but they have no analyst. Every “how did last week go?” or “which campaign is actually paying off?” became a request to a data analyst and a wait for a report.

What I built

1. An agent that takes the question in plain English, works out which data it needs, queries it and answers with the number and a chart made for that question.
2. Connectors for Shopify, GA4, Meta Ads and Google Ads, plus BigQuery and Snowflake for brands that already keep a warehouse.
3. Built for non-technical people: founders and marketers ask directly, without waiting for an analyst to build the report.

How it got into production

1. Built solo in one month, from the first line of code to the first customer brand using it.
2. Each brand connects its own stores and ad accounts; the agent only reads through read-only, allow-listed views.
3. Nine customer brands use it in production today.

The problem

People making AI video were juggling a separate app for every model, and paying for expensive canvas tools to tie them together. Every provider has its own API, its own rules for what you can feed it, and its own way of running jobs that take minutes. Friends of mine now use TwitCanva instead of paying for those tools.

What I built

1. An infinite canvas where every step is a node (prompt, image, video), joined by typed connections (text to image, image to video), so a whole shoot is one workflow you can see, rerun and share.
2. One backend that talks to six providers (OpenAI, Google Gemini and Veo, Kling, Hailuo, Fal.ai) and to local GPU models. API keys stay on the server; the browser never sees them.
3. A model picker that knows what each of the 13 video models accepts (text, an image, a first and last frame), so you can only connect what will actually run.
4. A LangGraph chat assistant that can see the images on the canvas and writes the prompt for the next step.

How people run it

1. Docker Compose for a one-command setup; setup scripts for Windows and Linux for the optional local GPU models.
2. Camera angle control deployed on Modal for people without a big GPU. It scales to zero after 5 idle minutes, so it costs nothing when no one is using it.
3. One-click publishing to X and TikTok through their OAuth flows, with a setup guide for each integration.

The problem

Agencies and ad tools get paid whether or not the ads work, so small merchants carry all the risk: they pay up front and hope. Bidtopus flips it. The merchant sets a target (say, 2.0× ROAS in 7 days) and the agent is paid only if it hits it.

What I built

1. An ML model underwrites every contract first: from the target, the window and the account’s history it estimates the chance of success (65.1% in the demo) and recommends accept, counter-offer or reject.
2. Claude negotiates the terms with the merchant in plain language, then drafts the campaign strategy.
3. Once the merchant approves, the agent runs the campaign on Meta Ads through MCP and checks the live numbers every 24 hours, with a forecast of where ROAS will land.
4. The fee is locked in USDC escrow on Arc when the contract starts, and released to the agent or refunded to the merchant when the window ends.

How it runs

1. Four parts, each deployed on its own: Next.js on Vercel, FastAPI and the agent on Railway, Postgres on Neon, the escrow contract on Arc testnet.
2. Circle Wallets hold the agent’s keys, so the agent never touches a raw private key.
3. Tests for security, idempotency and state gates, and evals for how the agent negotiates.