Twitter β’ LinkedIn β’ GitHub β’ Medium β’ Email
Iβm a Software Engineer with 7+ years in backend and mobile. Iβve shipped systems that reached over a billion users and ran services that process millions of messages a day. I care about open source, clear design, and reliable scale.
- Pion/WebRTCβbased media I/O library used by other services (egress/ingestion for transcription & recording; ingress for HLS, WHIP, SRT).
- Whisper infrastructure: language hot-reload config (no restart), Prometheus/CloudWatch metrics, Grafana dashboards, rollout speed-ups with Packer-baked AMIs, and CloudFormation maintenance.
- VAD with Silero via ONNX Runtime (Go) to enable real-time captions (no VAD β no live captions) and reduce waste.
- Accuracy work: built a hallucination dataset (transcribed a noise corpus) + Aho-Corasick matcher in Go, with deloops and Unicode range filters to remove hallucinations; important for most customers and critical for healthcare.
- Codegen & SDKs: internal OpenAPI β server-side SDK tooling; owner of the Python and Go SDKs (35k+ LOC each).
- Product support: outages on rotation, post-mortems, indexes for big imports, quick fixes across chat and dashboard (picked up Django fast when needed).
- p95 live transcription/closed captions: ~650 ms β ~300 ms.
- Cost: ~36Γ cheaper than OpenAI 4o transcribe for our load.
- Rollouts: ~11 min β ~1 min with Packer AMIs + trimmed Puppet.
- One service for transcription and captions β ~50% infra/ops reduction.
- Proved CPU wasnβt the bottleneck; eased GPU concurrency with NVIDIA MPS.
- Quality: sales team dogfoods our app as a Gong replacement.
Go β’ Rust β’ Python β’ C++/CUDA β’ Whisper β’ ONNX Runtime β’ Pion/WebRTC β’ AWS β’ CloudFormation (maintenance) β’ CloudWatch β’ Prometheus β’ Grafana β’ Packer/AMIs β’ Puppet β’ Kibana β’ Nsight β’ NVIDIA MPS β’ OpenAPI
- Medium: https://medium.com/@sachaarbonel
- Issues welcome: https://github.com/sachaarbonel/sachaarbonel/issues
- Email: sacha.arbonel@hotmail.fr