From Retell or Vapi to FreeSWITCH and Kamailio
What to move first, which call behaviors to keep, and how to compare an owned voice stack with the managed path.
TECHNICAL NOTES
Engineering voice systems,
in public.
Notes from building and debugging voice systems: failed calls, audio problems, architecture decisions and the evidence behind the fixes.
VOICE AIVOIPSIP / WEBRTCREAL-TIME MEDIADISTRIBUTED SYSTEMS
FEATURED ARTICLE
Debugging Audio Dropouts with Real Evidence
Ringback sounded stretched in the browser, but the recording was clean. Packet captures showed where 20 ms audio became 75 ms audio.
READ ARTICLE ↗Mean packet interval
- Before the fix
- 75.487 ms
- After the fix
- 20.003 ms
Packet captures and the timer fix inside.
LATEST NOTES
8 MORE ARTICLESREAD NOTE ↗
Voice AI Reliability Audit: What Breaks After the Demo
How I check latency, interruptions, answering-machine detection, call-state races and recovery in a Voice AI system.
A High-Load Telephony Platform Should Fail in Pieces
Where Kamailio, FreeSWITCH and RTPengine hold state, what limits their capacity, and which calls a failed node affects.
WebRTC Debugging, Layer by Layer
A debugging sequence from WSS and SIP registration to ICE, packet flow and browser playback.
One Call, Four Sources of Truth
Why our dialer needed separate counts for product calls and PBX channels, browser media evidence, and a clearly labelled AVMD review queue.
Designing Voicemail Drop Without Race Conditions
How we handed the customer leg to FreeSWITCH playback while releasing the agent, with retries and hangups accounted for.
What Actually Breaks at High Call Concurrency
What to test beyond active-call count: burst rate, media work, event lag, database contention and AI capacity.
State Machines Saved Our Dialer
Hangups, playback and retries kept racing with each other. An explicit call lifecycle gave every handler the same rules.