ImplementAI: The AI Implementation Engineer (Agora + Murf)
Link to open source: https://github.com/Rujalladhe/implementai.git
The problem
Every SaaS company pays engineers to sit on screen-share calls saying "click Settings... no, the other one." It's slow, mostly English-only, not there at 2 AM, and one engineer can take only one call at a time.
What ImplementAI does
A customer opens a meeting link and shares their screen. Our AI joins the call, reads the screen and guides them by voice, one action at a time. It notices every click and tab change, so the customer never has to say "done." When the steps are finished, it proves the fix with a real test. If it can't solve the problem, it brings a human into the same call with a Jira ticket.
The live demo
PizzaCo built a Hindi pizza-ordering voice bot on Agora with a Murf voice, but the bot can't understand Hindi and sounds American. Our AI reads the ticket and guides the customer through Agora Agent Studio and Murf's voice library: Hindi speech recognition, a Hindi FALCON voice and the right locale. It catches a wrong tab or a wrong click at once, warns about Murf's Gen2-vs-FALCON trap, and draws an arrow when the customer says "dikha ke batao." After Re-Publish, it starts the published bot on Agora in a private test channel to check that it really runs.
An Agora agent with a Murf voice fixes another Agora agent with a Murf voice.
How it's built
• Agora Conversational AI runs the call: speaker lock, so a TV or typing doesn't cut the AI off, and the interrupt API, so it stops talking the moment the customer does the step
• Murf Falcon is the voice: 16 languages, switchable mid-call ("can you speak Hindi?") in about 2 seconds, on the same step
• Claude is the brain, with tools that take real actions: run tests, search docs and past cases, create Jira tickets, ask experts on Slack or Teams
• Eyes: a Chrome extension plus Apple Vision OCR. A click is known in about 200 ms, and a screen is read in about 160 ms
• It learns: in Teach mode an expert does a task once and a LangGraph team of agents writes the guide, checking every step against the real screen. Every call ends with lessons learned
What makes it different
• Private by design: API keys, card numbers and passwords on screen are blurred before the AI ever sees them
• Helps before you complain: click the same button 3 times or hit the same error twice, and it offers help, already knowing the error
• Knows when to call a human: it searches Jira, Sentry, Grafana logs and past cases first, then a specialist joins the same call from their phone by scanning a QR code
• Works for any SaaS: upload your docs and connect GitHub and Jira, with no code change to your product
This build was uploaded as a hackathon project






