← All workCASE STUDY · WEBPHASE · FOUNDER-LED PRODUCT TEAM

A voice agent that joins the founder's meetings and briefs him out loud.

The founder wanted his briefing inside the calls he was already on, not in another dashboard. I worked directly with him to build a real-time agent that joins Google Meet, listens for its wake word and answers by voice.

ROLE: AI engineer, reporting to the founder
TIMELINE: Jun – Aug 2026
STACK: FastAPI, Recall.ai, DeepSeek, edge-tts, Supabase
ALSO DELIVERED: 3 audits for the CEO

THE BRIEF

Brief the founder where he already is: in the meeting.

A briefing is only useful if it reaches the person at the right moment. The agent had to join live calls, stay quiet until asked, answer from the company's own knowledge, and know when it was allowed to do more than talk.

WHAT MADE IT HARD

Real-time audio: replies have to land inside the flow of a live conversation.
The agent must not respond to itself or talk over people.
It needs clear limits on when it can only report and when it may act.
Its answers have to come from fresh company knowledge, every day.

HOW IT FITS TOGETHER

IN THE CALL

Google MeetLive call

Recall.ai botJoins and streams

FastAPI webhookVia Cloudflare Tunnel

Wake word?Otherwise stays quiet

DeepSeek + edge-ttsAnswers out loud

EVERY MORNING

Knowledge baseCompany docs

GitHub ActionsDaily TypeScript job

SupabaseStored and searchable

ClaudeWrites the briefing

WHEN IT MAY ACT

I designed a staged-autonomy model. Each step up gives the agent more freedom, and it only moves up when the founder trusts the step below.

STAGE 1EvidenceReports what the data shows

STAGE 2OpinionSays what it thinks it means

STAGE 3SuggestionProposes an action for approval

STAGE 4Autonomous actionActs on its own, within agreed limits

PRODUCTION FIXES

PROBLEMThe webhook handler blocked under loadFIXRemoved async/sync blocking in the handler

PROBLEMThe agent answered its own voiceFIXFixed a self-echo timing bug

PROBLEMRequests failed outside one regionFIXRemoved hardcoded region assumptions and wrong auth headers

PROBLEMA scheduled job fired twice, and a guardrail was brokenFIXCaught by reviewing live outputs, not job status

PROBLEMTwo system prompts disagreedFIXMerged into one versioned prompt

RESULTS

Beyond the agent, I delivered three audits to the CEO (business intelligence, product workflow and agent infrastructure) and set up agent loops on GitHub Issues for engineering work.

Liveinside Google Meet, answering out loud on its wake word

4autonomy stages, from evidence to autonomous action

DailyClaude-written briefings from the company knowledge base

NEXT CASE STUDYWhatsApp sales agent for Rizm TechRead →

Need an agent you can trust? Let's talk it through.

Book a free call WhatsApp me