.png&w=256&q=75)
1
1
Germany
1 year of experience
AI automation developer based in Berlin, originally from Ukraine. I turn manual workflows into automated systems with Python, LLM APIs and speech/media tooling. - Built and run an 18-step AI video localization pipeline for a real German-market client (Whisper, Claude, GPT, ElevenLabs/VMEG, FFmpeg, Playwright) with a SQLite job queue, automatic recovery and 511 tests - Built a multi-agent AI video production backend (FastAPI, PostgreSQL, Redis/Celery, Docker) - KeepAlive (AssemblyAI Voice Agent Hackathon): owned the voice/audio layer, including a FastAPI streaming TTS service, a Web Audio engine with priority queues and ducking, safety filtering of LLM text before speech, and Whisper-based speech verification Languages: Ukrainian, Russian, German, English. Open to AI automation projects and junior roles.

## The Problem Every year over 350,000 Americans suffer out-of-hospital cardiac arrest. Survival depends on the first critical minutes, but during CPR the rescuer's hands are physically locked on the sternum, pressing twice a second. Swiping a phone stops compressions. KeepAlive eliminates hands from the equation: the phone sits on the ground, and your voice becomes the interface. ## The Resuscitation Flow When a bystander shouts *"Help! My dad collapsed, he's not breathing!"*, AssemblyAI Universal-3.5 Pro streaming transcribes the panic in real time. A sub-millisecond local router locks the AHA-BLS protocol, speaks the first command, and starts a Web Audio 110 BPM metronome with visual depth guidance. When the rescuer asks *"Did I break a rib?"*, the answer arrives over the beat without pausing rhythm. Once paramedics arrive, the timers freeze and an EMS handover record (NEMSIS v3.5 shape) is read aloud and generated as a QR code for paramedic tablets. ## Two Brains, One Safety Boundary - **Deterministic Engine:** Controls all medical execution via an AHA-BLS state machine and pre-rendered audio. No generative model can alter CPR instructions. - **Generative Companion:** Answers panic doubts via Groq LPU behind a deterministic safety gate: max 18 words during CPR, zero dosage advice, and a mandatory return-to-beat directive. ## Measured Benchmarks & Stack - **Triage routing:** 0.49 ms | **Groq companion:** 0.38 s | **92 passing tests** - **Unified Voice:** 58 broadcast-mastered ElevenLabs clips keyed by asset ID, with dynamic fallback via `/speak` in the identical voice. - **Stack:** AssemblyAI Universal-3.5 Pro Streaming (16 kHz PCM WebSocket), FastAPI, React/Vite, Groq LPU, ElevenLabs, Web Audio API, Render. *Note: 911 CAD dispatch is simulated (MOCK). EMS handover follows NEMSIS v3.5 schema. Hackathon prototype under MIT license.*
30 Sep 2026