One lab doesn’t get sellers ready for a technical customer conversation. A program can, without pulling engineers off the roadmap to do it. This page is how I’d build one: what to find out first, the labs, how a seller is certified, and what gets measured.
A design by Tony Morinello, not LiveKit’s plan. It was written without seeing LiveKit’s deals, its sellers or its training, so its first step is to go and look. Only Lab 1 exists. Everything marked Proposed is a plan.
Before anything else
Find where deals stall
The program starts with listening. A curriculum written before that teaches what its author finds interesting, which is rarely what loses deals.
What comes out is a ranked list of gaps. It sets the order of the labs below, and I’d expect it to change that order.
Twenty recorded calls
Ten won and ten lost, picked by the CRO. I’d mark every moment a seller dodged a technical question, guessed at one, or said “let me get an engineer.”
The engineers who get pulled in
Which questions do they answer every week? Each repeat is a seller card that hasn’t been written yet.
The check, for everyone
Every seller takes Lab 1’s five questions cold, plus a handful per planned lab. Two minutes each, and it’s the baseline the rest is measured against.
Lost-deal notes
Where the reason given is “technical fit” or a competitor’s name, what was the question behind it?
The curriculum
Six labs, each answering something a customer says
The labs cover four areas: LiveKit fundamentals, voice AI architecture, agents and the competitive landscape. Each lab starts from a sentence a seller will hear on a call and ends with the question to ask back.
The last one is the hardest to write from outside: the named comparisons need LiveKit’s product team behind them, and a seller should leave it able to say where LiveKit is the wrong choice.
A caller’s signal drops to one bar. Which call falls apart first, and how does it sound?
Hear it
The same call on a good connection and on one that loses a packet in every ten.
Try it
Join one room from your laptop and your phone. Mute one, walk away from the Wi-Fi, and watch what the other sees.
Ask this
“What happens to a call today when the caller is on a train, or on one bar of signal?”
Check
Rooms, participants and tracks, in the words a customer would use.
Practice call
A prospect whose users are field technicians on patchy mobile data.
Lab 3: From a phone number to an agent
Proposed
Predict
A caller dials a number. What has to sit between the phone network and the agent?
Hear it
Juniper answering the same booking from a browser and from a real phone line.
Try it
Call Juniper from your own phone, then find your call in the project’s sessions.
Ask this
“How do callers reach you today, and who carries those calls?”
Check
What a SIP trunk is for, and which side of it LiveKit runs.
Practice call
A clinic group moving its appointment line off an old phone menu.
Lab 4: Build Juniper yourself
Proposed
Predict
You give the agent a tool that looks up opening hours. When does it use it, and when does it guess?
Hear it
The same question put to an agent with the tool and to one without.
Try it
Build the restaurant agent in Agent Builder, in a LiveKit Cloud project of your own, and give it one tool.
Ask this
“Who would build the first version on your side? What would it need to look up or do?”
Check
Instructions, tools and models: which one you change to fix which problem.
Practice call
A founder who wants a working demo by Friday and has one engineer.
Lab 5: Finding out what went wrong
Proposed
Predict
A caller waited four seconds for an answer. Where do you look first?
Hear it
One bad call, replayed with its transcript and the timing of every step.
Try it
Open one of your own Lab 1 calls in Agent Insights and find its slowest turn.
Ask this
“When a call goes badly today, how do you find out what happened?”
Check
Reading a turn’s timing, and telling a slow model from a slow hand-off.
Practice call
A head of support with a complaint, a call ID and no engineer in the room.
Lab 6: Build, buy or bolt on
Proposed
Predict
A prototype works at a thousand minutes a month. What breaks first at a million?
Hear it
One customer’s situation, argued three ways: a hosted platform, a model provider’s own API, and building it all in-house.
Try it
Take that customer through each choice and say, out loud, where LiveKit is the wrong one.
Ask this
“What’s the prototype built on, and what’s the first thing that breaks as you grow?”
Check
Which kind of alternative a prospect is describing, from two sentences.
Practice call
Priya again, six months on, with a competitor’s quote in hand.
Who takes what
Two paths through the same labs
A Solutions Architect who has only read the code explains a dropped turn in terms of thresholds. One who has also sat through the seller’s labs explains it in terms of a caller reading out an account number. I want the second kind.
AEs and SDRs
Hold a technical conversation with a customer, and know when to bring in an engineer.
The six labs, one or two a week. Lab 1 takes about half an hour.
A practice call after each pair of labs, with a new prospect each time.
Certified for discovery on a passed practice call.
Presales and Solutions Architects
Everything a seller learns, and then the code behind it.
The same six labs first. They need the customer’s words as much as anyone.
Then they change Juniper: add a tool, swap a model, read the agent’s own timing. How it’s built is the first unit.
Certified for demos by running one live and changing a setting mid-call.
Certification
Ready means you’ve done it out loud
A quiz can’t certify this. A seller who picks the right letter can still freeze when a CTO asks whether they could self-host.
So both certifications are things done out loud, and both are scored the same way: a point needs a moment in the recording that earned it.
Discovery: the practice call
Built
Eight to ten minutes with an AI prospect. A pass is 8 of 12 points with no made-up answers; one bluff fails the call, whatever the total.
The seller runs Juniper for a colleague playing the customer, switches one setting in the middle of the call, and says what changed and why that customer should care. The lab already switches settings on a live call, so the demo needs a script and a rubric, not new software.
Every release
Keeping it true after LiveKit ships
Training goes stale quietly. LiveKit’s docs now mark its text-based turn detector as deprecated in favour of the audio one; a lab written a year ago would still be teaching the old one, and nobody would have been told.
Lab 1 is built so that going stale makes a test fail.
Every product fact has a source and a date
Each seller card lists the LiveKit docs pages it rests on and the day they were last read. A test fails if a card has no source.
A release is read against the cards
The changelog is checked against that list of pages. A card whose page changed gets re-read, and re-dated or rewritten.
Recordings are re-made, not trusted
Each experiment is recorded from a real call in a few minutes. A test fails if a recording no longer shows its lesson, such as a baseline that stops interrupting.
The agent and the pages can’t drift apart
The names they share (the RPCs, the scoring rubric, the facts the judge marks against) are compared by a test on every push.
What gets measured
Four measures, and what Lab 1 records today
Two of the four can be read off the lab now. The other two need a sales team’s own systems, and I’d rather say so than pass off a stand-in as the real thing.
So far the only data is one run of my own: 2 of 5 on the check before the lab, 5 of 5 after. That shows the recording works and nothing else. A pilot with two or three people who don’t know LiveKit comes next.
Measure
Recorded today
Still needed
Seller confidence
Three ratings from 1 to 5, before the lab and again after the practice call.
The same ratings a month on, once the seller has taken real calls.
Assessment pass rates
The check’s score before and after. Each practice call’s score on six items, and whether it passed. Whether each prediction was right.
How many tries a pass takes, across a whole group of new sellers.
Time to first qualified technical conversation
A stand-in: the time from starting the lab to the first passed practice call.
The real number is in the CRM: the first meeting a Solutions Architect accepts as qualified. What “qualified” means gets agreed with the CRO before anything is counted.
Ramp time
When each step was started, so how long each part takes.
A start date and a first closed deal, which live in the HR system and the CRM.