Custom AI agents, workflow automations, and AI-powered web apps for founders and small teams. I'm Igor Lima, a former operator who learned to build the tools I wished I'd had.
Everything below is a real, scoped engagement. Prices are starting points based on typical complexity. We talk, I send a fixed quote, you say go.
An agent that handles one workflow end to end. Built around your business, not a generic chatbot wrapper.
The thing your team does 40 times a week, automated so it takes 30 seconds and zero attention.
A small, sharp web app with AI doing the heavy lift. Real product, not a sandbox demo.
Move your messy data through a language model, structure the output, drop it where the rest of your stack expects it.
Fixed scope, fixed price, working demo in seven days. For pitching investors or pressure-testing an idea before you sink real money in.
Drop me into a working session. We sketch the architecture, write the hard part together, you leave with something you can ship.
Three live products you can try right now, more on the way. Want something like these built for your business? That's the offer.
427 career-grade prompts built on a 9-layer framework (Mission, Agent, Situation, Task, Examples, Rules, Think, Eval, Output). Templates, completed exercises against real situations, and a scheduled scraper that surfaces the best prompts from across the web.
Most prompt libraries online are unsorted piles of clever-sounding strings. When a prompt fails, there's no framework to tell you which layer went wrong, and "just rewrite it" is not a methodology.
Every prompt is built on the MASTER framework so any failure traces to a specific missing layer. 427 prompts across 18 categories, each one audited. Two formats: Full MASTER (all 9 layers) and MASTER-lite for daily-use templates.
Next.js static export → GitHub Pages, plus a scheduled scraper that scores and ingests prompts from across the web into the discovery feed.
Backs up the "prompt engineering since 2020" claim with a clickable, browsable artifact. The framework itself is the deliverable that hiring managers and prospects actually want to see.
Browser-native statistical analysis. Drop in a CSV, get descriptives, distributions, and correlations rendered in the page. Pyodide runs the Python in WebAssembly so your data never leaves your machine.
Quick CSV analysis usually means uploading to a server you don't fully trust or wrestling with a local Python install. Neither is great when the data is sensitive.
Run pandas in the browser via Pyodide (Python compiled to WebAssembly). The CSV never leaves the device, but the full pandas API is available.
Vite + React UI → Pyodide worker → pandas, NumPy, SciPy in WebAssembly
Sensitive data stays local. No backend means zero hosting cost. The whole site is a static bundle on GitHub Pages.
Image-to-video AI generator in the browser. Drop a photo, write a short motion prompt, get a 5-second MP4 back. Built on Wan 2.2, the open-source answer to Runway Gen-3 and Kling.
Image-to-video AI is locked behind paid platforms (Runway, Kling, Pika). Hobbyists and small teams who want to try it hit paywalls or limited free tiers.
Proxy a public Hugging Face Space running Wan 2.2 through a Cloudflare Worker. Static frontend on GitHub Pages calls the Worker, never the Space directly.
Astro static site → Cloudflare Worker (CORS + URL stability) → HF Space (Wan 2.2 on shared GPU)
Worker abstraction means swapping inference providers (fal.ai, Replicate, private Space) doesn't touch the frontend.
Updated whenever something ships or moves. If you're hiring me, this is what active looks like.
Intent classification + reply drafting for inbound store leads
Streaming progress UI, last-frame guidance, longer outputs
I'm a 2nd-year CS student, but I came to the major sideways. Before college I owned and operated a small chain of mattress stores. Payroll, inventory, leads, the whole thing. That's where I learned that the gap between a decent business and a great one is usually how much busywork the owner is drowning in. Optimization stuck with me.
I started prompt engineering in early 2020, back when "AI tools" meant Google Colab notebooks and NightCafe for image generation. The proper code came later, in the last couple of years. So my hands-on time with language models is longer than my time as a "developer," which is rarer than it sounds.
I'm trilingual: English, Portuguese, Spanish, all fluent. I grew up in Brazil and have been a US citizen since 2011. If your customers or operations span Latin America, the US, or Brazil, that's a moat I can lend you.
A short note is enough to get started. I reply within one business day. If your project's a fit, the next step is a 30-minute call to scope it and lock a price.
