Artem Techman | Founder, Ryon.ai How We Built a Call Center Solution Processing 50,000 Calls a Day Call center agents spend 40 percent of their shift on post call work. They type notes into CRMs. They summarize conversations. They extract customer details like order numbers or complaints. One client processed 30,000 calls daily. Agents handled […]
Artem Techman | Founder, Ryon.ai
How We Built a Call Center Solution Processing 50,000 Calls a Day
Call center agents spend 40 percent of their shift on post call work.
They type notes into CRMs. They summarize conversations. They extract customer details like order numbers or complaints.
One client processed 30,000 calls daily. Agents handled 45 calls each per shift before admin ate the rest. We fixed that.
What We Built
Ryon.ai Call Center Summariser sits as middleware. It connects telephony systems to CRMs.
Calls hit our API endpoint. Audio streams in real time via WebSocket. We transcribe with OpenAI Whisper large v3. Accuracy hits 92 percent on clean audio.
Next, we feed transcripts to GPT 4o for summarization. It outputs a 150 word summary. Plus structured JSON: customer name, intent, sentiment score from 1 to 10, action items.
Data posts to the CRM via webhook. Agents see it in 45 seconds post call. No manual entry.
We support 150 languages. French, Mandarin, Arabic. Detection runs first, then language specific transcription.
The Architecture
Full replacement kills adoption. We built middleware instead. It plugs on top of existing telephony and CRM.
Telephony like Twilio or Genesys sends audio to our ingress gateway on AWS ALB. We use Kubernetes clusters across three regions: eu west 1, us east 1, ap southeast 1.
Ingress fans out to 200 transcription pods. Each pod handles 250 concurrent streams. We queue with Apache Kafka at 100 MBps throughput.
Post processing: 50 summarization workers on GPU instances. GPT calls batch in groups of 20. Latency stays under 30 seconds for 95 percent of calls.
Output routes to CRM APIs. Salesforce, HubSpot, Zendesk. We maintain 50 connectors. Authentication via OAuth 2.0 or API keys.
Scale hits 50,000 calls daily now. Average call 4.2 minutes. Peak at 4,000 calls per hour. We auto scale pods from 50 to 500 in 90 seconds.
The middleware approach means clients can swap telephony vendors. Our layer stays. Zero downtime migrations.
The Results
Before: Agents spent 12 minutes per call on admin. Total shift 480 minutes. Calls per agent: 32.
After: Admin drops to 1.2 minutes. Calls per agent: 38. That is 20 percent more volume.
One client went from 30,000 to 36,000 calls daily. Same 950 agents. Manual summarization time fell 90 percent. From 2 hours to 12 minutes per shift.
Error rate on data extraction: 3 percent now. Was 18 percent with manual entry. Sentiment scores match human labels 87 percent of the time.
Cost: /bin/sh.12 per call processed. Telephony billed separately. ROI in 14 days for most clients.
What Surprised Us
Noise killed transcription first. 22 percent of calls had background chatter. Accuracy dropped to 71 percent.
We added RNNoise suppression. Pre processes audio in 200 ms. Accuracy back to 91 percent. It costs 15 percent more CPU but the trade is worth it.
Second surprise: agents wanted more. Week one usage was 82 percent of all calls. They started requesting custom fields like an upsell opportunity score. We shipped that in sprint two. Adoption beat our 50 percent target.
Three Lessons for Anyone Building Similar
Lesson one: Integration rules all. Spend 40 percent of dev time on connectors. We wrote 50. Clients can test in 2 hours.
Lesson two: Real time feels better than it is. Live transcription feels instant. Summaries can batch for cost. Our GPUs run at 70 percent utilization and we saved 35 percent on inference bills by batching.
Lesson three: Monitor drift hard. Language models shift. We A/B test weekly on 1,000 gold standard calls. If accuracy dips 2 percent we retrain the prompts.
Where This Is Going
We process 50,000 calls today. Target is 200,000 by Q4 2026.
Next up: voice biometrics to detect fraud in 1.2 seconds. Then real time agent assist during the call itself, with prompts delivered via earpiece. Early tests show 15 percent conversion lift.
Ryon.ai builds intelligence layers for large companies. This solves one pain point. Many more to come.
Artem Techman is the founder of Ryon.ai, an AI integration agency for large corporates based in Cannes, France.
/ Ready to see where AI creates leverage in your business?