We listen to the calls nobody else can
Voice agents fail quietly. The transcript reads fine, the latency graph stays flat, and the caller hangs up confused. The only way to know what your agent actually said is infrastructure that grades every call, and that is what we build.
Why Sonora exists
We got burned first
In early 2025 we were two engineers running a scheduling agent for dental clinics. It booked thousands of appointments a week. One Tuesday we swapped the underlying model to cut latency, ran our usual smoke tests, and shipped.
The swap broke number confirmation. In one out of five calls the agent stopped reading the appointment time back to the caller, so patients showed up on the wrong day. Every dashboard stayed green for nine days. We found out when our largest clinic churned and mailed us the recordings themselves.
That week we built the tool we needed: replay real calls against every change, grade each turn, page a human the moment a number moves. The internal script became Sonora. Now it grades other teams' agents, so they hear the bad call before their customers do.
- Calls graded each month
- 500,000+
- Voice teams on Sonora
- 120
- Days our own failure went unnoticed
- 9
How we work
Four rules that decide what ships and what waits.
Listen to the tape
Decisions come from real recordings. When a dashboard and the audio disagree, the audio wins.
Precision over polish
A scorer that is exactly right about digits, dates, and names beats a prettier chart every time.
The customer hears everything
A half-second stall or one wrong digit never shows up in logs, so we grade the call the way the caller heard it.
Boring reliability
Evaluation infrastructure has to outlive the systems it watches, so we build for determinism, uptime, and dull on-call weeks.
The team
Five people who have shipped voice products and carry pagers for this one.
Founding team.
Dr. Maya Okafor
Co-founder & CEO
Maya spent seven years in speech ML, working on ASR robustness and speech evaluation before Sonora. She has published on turn-level dialogue evaluation and still reviews a sample of graded calls every Friday.
Ben Whitfield
Co-founder & CTO
Ben ran observability systems handling millions of events per second at a logistics company. He built the grading pipeline to the same standard: deterministic, replayable, and quiet at 3 a.m.
Rachel Donnelly
Founding Engineer
Rachel built the scenario runner and the CI deploy gate that blocks a bad prompt before rollout. Before Sonora she replaced IVR systems at two regional banks.
Marcus Bell
ML Engineer
Marcus owns the scoring models, including the digit, date, and name verification stack. He previously trained speech models for an in-car assistant shipped across 40 vehicle lines.
Tyler Nguyen
Developer Experience
Tyler writes the SDKs, the docs, and the error messages you actually read. He came from a CPaaS company where he supported 2,000 developers building call flows.
Security
Your calls contain names, numbers, and health details
We handle recordings the way you would want your own handled. Every control below is live today, not on a roadmap.
SOC 2 Type II
Audited annually against the full trust criteria. The current report is available under NDA.
Encrypted in transit and at rest
TLS 1.3 on every connection, AES-256 on every stored recording and transcript, keys rotated automatically.
No training on customer data
Your audio and transcripts grade your agents. They never train our models or anyone else's.
Retention you control
Set retention per project, from zero retention (grade in memory, store only scores) up to 7 years. Change it any time without a support ticket.
Deletion that sticks
Hard deletes via API or dashboard, propagated to backups within 30 days. A DPA with SCCs and data residency options come with the Scale plan.
Work on this
We hire engineers who have shipped voice products and can explain exactly what broke. Small team, real calls, production stakes from week one.
Get in touch