Samuel Osei, homeSay hi
Work

Sika

Sika helps people invest on the Ghana Stock Exchange by chat, in eight languages.

  • AI product
  • 2026
  • Solo build · Full-stack and AI engineering
Screenshot of the Sika homepage
languages, including Twi, Ga, Ewe and Hausa
8languages, including Twi, Ga, Ewe and Hausa
red-team attacks stopped
14/14red-team attacks stopped
days from first commit to submission
7days from first commit to submission

The brief

Most people in Ghana who could invest on the stock exchange never start. The tools assume you already know the jargon, read English comfortably and want another app. I built Sika for the Meta Global AI Developer Hackathon, whose theme was "Build for Your World": an investing assistant you talk to in WhatsApp, in your own language, by text or voice note.

What it does

  1. Five questions build your risk profile.
  2. You make a plan together: a goal, a monthly amount and your limits.
  3. Sika suggests trades that fit the plan, priced from the live order book, with every charge shown.
  4. You place the order with your own broker and tap "I bought it". Sika never trades for you.

Around that it answers questions about the market today, single stocks and your portfolio, compares stocks, and sends sourced news about the ones you follow.

How I built it

The rule I built around is the AI suggests, code checks, you decide.

  • The agent is a tool-calling loop on Meta's Muse Spark model. It can look up prices, read the plan and draft a suggestion, but it can't send one on its own.
  • A rules engine checks every suggestion against the investor's plan and limits in ordinary code. If it breaks a rule, it doesn't go out.
  • A fee engine prices each trade to the pesewa, so the cost shown is the cost paid.
  • Translation is fact-checked. Replies are translated into Twi, Fante, Ga, Ewe, Hausa, Dagbani or Pidgin. If translation changes a number or a stock ticker in a line, that line is sent in English instead.
  • Voice works both ways. WhatsApp voice notes are Ogg Opus, which I decode and encode with WebAssembly inside the Worker.
  • A guardrail keeps it on topic. In red-team testing it stopped 14 of 14 attacks, including attempts to make it ignore the investor's limits.

Where it runs

Everything runs on Cloudflare's free tier. A Worker built with Hono receives WhatsApp webhooks and puts them on a Queue, so Meta gets a fast reply and the slow AI work happens after. Data lives in D1. Scheduled jobs record the market every 30 minutes and send news on weekday mornings and afternoons.

Status

Sika is a hackathon build. The web demo is open to everyone, the code is on GitHub, and there is a demo video. The WhatsApp number is still a Meta test number, so it only answers verified phones.