A real production call — 3 minutes 10 seconds, recorded 2026-04-13 — replayed alongside the systems that ran it. Press play, watch the transcript light up, and see every knowledge-base retrieval and tool call land as they happened.
Call ID
call_3c900a9c…b963ed
Agent
Aria — Sendsteps SDR
Version
v7
Channel
web_call
Recorded
2026-04-13 19:17 CEST
Duration
3m 10s
Outcome
meeting_booked
Tuesday April 14 · 11:00
Sentiment
Positive
Interaction Scorer · Paperclip
KB retrievals
8
inline, during call
Tool calls
3
check_availability · book_meeting · end_call
End-to-end latency
3017ms
cost $1.071 · 11,991 tokens
Listen to the call
0:00 / 3:10
Audio streams from Retell's CloudFront. The transcript and event log below stay in sync — click any line to jump.
What's happening right now
Live commentary
0:00 — Pre-call
Press play to begin. As the call unfolds, this box explains the system-level move Aria is making at each moment — the qualifying questions, the KB lookup, the moment she invokes a tool, and what each one writes back into Zoho.
Transcript
Aria ↔ Pete · click any line to seek
System events
KB · tools · disposition
After Aria hangs up — what the backend did
The call ends at 3:10. Within seconds, the rest of the Pillar 2 stack picks up the work: the retell-webhook-receiver service captures the call_ended event, persists the transcript and the structured outputs to PostgreSQL, fires the Interaction Scorer on the transcript, and routes the meeting confirmation through to TidyCal and Zoho. Here is the post-call data exactly as it landed in our systems.
Paperclip · Interaction Scorer output
call_outcome
meeting_booked
call_disposition
meeting_booked
user_sentiment
Positive
call_successful
Successful
prospect_email
COELEWIJP@gmail.com corrected mid-call by Pete — original guess included a dot
objections_raised
Concern about the cost of the service. Handled at 1:31 with the free-pilot framing and the €75/teacher/year price point.
qualification_signals
Uses interactive tools such as Mentimeter and Kahoot; interested in post-lecture AI analysis to identify student struggles; values time-saving features for lecture preparation.
Auto-generated summary
The agent contacted Pete, who appears to be involved in university lectures, to discuss Sendsteps' AI tool for enhancing student engagement and post-lecture insights. Pete currently uses tools like Mentimeter and Kahoot but showed interest in the AI analysis feature despite initial cost concerns. The call concluded successfully with Pete booking a demo meeting with a specialist named Mike for April 14th at 11:00 AM.
Tools fired
check_availabilitybook_meetingend_call
KB retrievals
8 inlineacross the call — pricing, product features, scheduling logic
End reason
Agent_hangup Aria closed the call after confirming the booking — clean exit, not user-initiated.
The systems behind one call, in five stages
Every minute of audio you just heard touched five layers of our stack. Here is the same call described in system-of-record terms.
01 · Retell
Voice + LLM loop
Aria's audio runs in Retell — STT, TTS, turn-taking. The agent itself is configured as a Conversation Flow (agent_aa56b6…29b, v7) backed by our system prompt.
02 · Our backend
retell-webhook-receiver
A single Node service on Railway ("gracious-unity / production") catches Retell's tool-call requests — check_availability, book_meeting — and runs the actual business logic.
03 · Booking
TidyCal / Calendly
tidycal.js queries live availability for Mike de Vries during the call, returns three slots, and books the chosen one before Aria says "scheduling that for…".
04 · CRM
Zoho upsert
Lead record created/updated with Pete's email, qualification signals, objections, and meeting time. Same path every Aria call writes through.
05 · Paperclip
Interaction Scorer
Post-call, the Interaction Scorer (interaction-scorer.js, running Claude) reads the transcript from PostgreSQL and produces the analysis you see above. This is the safety net for every call.
Paperclip · the observability layer
The agents are watched. That's how we sleep at night, and how the pilot scales.
Paperclip is the platform that watches every other agent in our stack. Without it, we'd be deploying LLM-driven systems into customer-facing pilots on faith. With it, every call Aria runs and every chat VentoBot will run produces a structured, auditable record — and a set of dedicated agents whose only job is to grade those records, reconcile them against external systems, and alert us when something drifts.
It's a separate Railway deployment from the retell-webhook-receiver service that runs Aria. Same PostgreSQL, separate process, separate UI at paperclip-production-d112.up.railway.app. As of today it costs us $4.89 / month to run, and it has produced 80 scored interactions, a daily reconciliation heartbeat, and four weeks of intelligence reports.
How it works
Conversation events arrive in PostgreSQL via webhooks. Six dedicated agents read from that shared store and write back two things: findings into the dashboard, and alerts/digests into email.
Interaction Scorer
● Live
Scores every call transcript against a rubric — tone, accuracy, on-script, refusal handling. 80 calls graded to date (50 NL, 28 EN, 2 EW). Implemented as interaction-scorer.js, running Claude.
Reconciliation Checker
● Live
Daily 06:00 CEST. Counts Retell webhook events vs PostgreSQL records. Fires high-priority emails when they don’t match within 2%. Catches silent ingestion failures.
Chief of Staff
● Live
The orchestrator. Coordinates work across the other agents and writes most of the human-readable summaries that show up in the dashboard’s Recent Tasks panel and the weekly emails.
Intelligence Analyst
● Partial
Produces weekly intelligence reports (latest: SCA-30, week of Apr 23–30). Not yet emitting live latency / drop-off / failure-rate metrics — that’s the partial bit.
CEO
● Paused
Currently paused on the dashboard. In the inventory but not generating output today.
Pipeline Sentinel
● Blocked
Designed to watch pipeline health. Blocked since first heartbeat ~1 month ago on missing credentials (issue SCA-2). Has never run successfully. Either unblock or retire.
What Paperclip actually sends you
Three real templates — the high-priority alert, the daily heartbeat summary, and the weekly intelligence digest. Content reconstructed from the live dashboard’s send log.
inbox · alerts@scala-auxilio · 2026-05-04 06:04 CEST
What I think happened: 21 chat_ended events on 2026-05-04 arrived with agent_id=null and call_id=null. That’s 80% of all-time null-id events in a single day. Most likely cause: the EconoWind webhook config is not yet fully wired — expected during the controlled pre-launch hold pending EconoWind’s approvals.
Open issue: SCA-36 — pending Retell API access for cross-validation (currently webhook-only). When VentoBot goes live this week and the webhook config completes, this should clear automatically.
— Reconciliation Checker (Paperclip) View in dashboard: paperclip-production-d112.up.railway.app/SCA/issues/SCA-36
inbox · alerts@scala-auxilio · 2026-05-06 06:02 CEST
From: Chief of Staff (Paperclip) <reports@scala-auxilio.com> To: petrusc@adsum-auxilio.com Subject tag: [SA Weekly] · task: SCA-30
Compiled by the Intelligence Analyst, summarised by Chief of Staff. Read time: ~3 minutes.
Headline numbers (week of Apr 23–30):
Aria EN: 28 outbound calls placed · 4 meetings booked (14% booking rate) · positive sentiment on 71% · objections led by “cost” (44%) and “already use Mentimeter / Kahoot” (33%)
Aria NL: 0 calls (paused — Dutch-LLM quality, open investigation)
EconoWind Chat: sandbox-only this week; integration test traffic only
Reconciliation: 6 of 7 days CLEAN; one mismatch flagged 2026-04-25, traced to a brief Retell webhook retry storm — auto-cleared next morning
Top quality flag from the Interaction Scorer: on three EN calls Aria over-anchored on the Utrecht Medical Centre case study even when the prospect was clearly in a non-medical faculty. Adding faculty-aware case-study selection to the KB retrieval is queued as a prompt update — assigned, not yet shipped.
Recommended action for next week: ship the faculty-aware case-study selection (low-risk, fast); decide on the Pipeline Sentinel credentials so we can either unblock or retire it before the CTO walkthrough.
— Chief of Staff (Paperclip) · full report: paperclip-production-d112.up.railway.app/SCA/tasks/SCA-30