Liva Founding Operations Lead Working Demo

The model problem

A voice model interrupts people while they're still thinking.

The fix isn't more data. It's data that keeps the pauses, overlap, and repair the model is missing, without letting broken captures through.

The demo follows one real conversation I recorded through the full operating path. It shows how I would define the work, decide what code can check, route the judgment calls, and use each review to improve the next batch.

Nothing here touches a real Liva engagement. The conversation is real. The batch around it is invented to show the controls.

The customer signal

The complaint is "get more natural conversations." That's not a brief. It doesn't say what behavior is missing, who can produce it, what to return, or what needs a decision from the customer.

So the brief for this case names all four. Collect: a real decision between two people who know each other, pauses and backchannels and talk-over and repairs left intact. Return: capture faults, missing metadata, unusable rights records. Escalate: anywhere the customer's goal and the acceptance rule disagree.

Listen before reading the rule. The hesitation the model interrupts is right there in the recording.

Real, consented input

The model interrupts during natural hesitation. Listen in context before deciding what the rule should do.

0:00 / 2:52
Open the checked, unattributed transcript

Produced locally with an open-source model, then checked against the audio. Turns stay unattributed because a mono recording cannot establish who said what.

    Staff it before anyone records

    Every handoff needs an owner, an output, and a next move. Match the work to the people first.

    Operations translates the signal into the brief and the quality rules, staffs contributors and reviewers, sets deadlines, watches blockers, and sends open questions and delivery risk to the customer. Contributors confirm consent, record with someone they know, run a ten-second level check, then leave the pauses and repairs alone. Reviewers with the right language and context check the annotations against the current instructions. They make the calls the automatic checks cannot, record why, and explain exactly what to fix when a session has to go back.

    What code owns, what people own

    Some checks establish facts. AI can prepare the review. Meaning and release stay with a person who can defend the call.

    Deterministic checks

    Code establishes what happened.

    Format, clipping, level, dead air, missing fields, duplicate files, and rights holds each produce a measurement, a limit, and one exact return instruction. No interpretation.

    AI-assisted preparation

    AI prepares the review.

    In a fuller workflow it can draft transcripts, suggest timestamps and event tags, and rank likely problems. All of it is work to verify, not ground truth.

    Qualified reviewer

    A person owns context.

    Hesitation, repair, channel realism, and cultural and language nuance need the right reviewer. The output is a decision with a reason, an owner, and something the operation learns.

    If the same false positive keeps coming back, the operation should learn from it. It can become a reviewed example, a rule change, or a clearer instruction.

    Three moments show where the line sits. Two are from the consented conversation. The third is a copy I deliberately damaged.

    Control the batch

    The recording above is one real 2:52 sample. I placed it beside twelve invented sessions to show the controls against an 8-hour delivery goal. The lengths, rates, dates, and costs are made up. What's real is the shape of the tracking.

    Right now: 4.20 hours ready, 3.80 still needed, 66.9% accepted without a rerecord, and four sessions needing action.

    Those four are the whole point. A low recording level, returned, waiting on a level measurement and a re-record deadline. Phone-channel hum that fails the current rule, returned, and now a question for the customer about whether realistic phone channels should route to review. A consent record that isn't linked, blocked until the rights are resolved. And a natural pause that tripped the dropout check, in review, where a senior reviewer decides in context and saves the example for future reviewers.

    Before the next batch, I would make three changes. I would save the reviewed pause as an example for future reviewers. I would send the clipping measurement with one re-recording instruction. And I would take the recurring channel question to the customer instead of quietly changing the rule.

    What I excel at

    I take an unclear request, turn it into an operating plan, build the system under it, and stay responsible for what ships.

    Consulting now, I turn ambiguous client work into documented SQL, Python, low-code, and AI-assisted systems, then review the outputs for edge cases and release readiness. Before that I ran operations for 30+ time-sensitive programs, supervised volunteers and interns, and managed contractors and trainers. On a small team I built the SQL and Python workflow that handles thousands of closely monitored certificates and has saved hundreds of staff hours.

    The bench

    Here is the working version.

    Start with the phone-hum file. The check measures the signal, the customer's goal changes who decides, and the human decision updates delivery and the next batch. The other test files, full batch, and math are still here if you want to inspect them.

    The consent record for the real clip: verified, scoped to a public demo plus altered quality-check copies, sale and model training prohibited, asset hash shown, consent audio held privately and never shipped with this site. Kris agreed to that on tape. The private source and the consent recording stay off the page. The public clip and every test file record where they came from and how they changed.

    What the automatic checks enforce

    CriterionRule

    Permission and where the audio came from (Provenance)

    Where the audio came from

    I kept the original recordings private. I checked that the public clip contains the conversation, not the separate consent take, and made each quality-check file as a labeled alteration of one excerpt.

      See the technical record

      The complete record is also available in audio provenance.

      Test what code can decide

      Start with one file. The check can prove the hum is above the limit. It cannot tell us whether that sound is useful phone realism or a broken setup.

      Working example: phone-line hum

      The number is clear. The meaning isn't.

      Measured
      52.2 dB hum contrast
      Starting limit
      12 dB
      What code knows
      The file contains persistent hum above the limit.
      What code cannot know
      Whether the customer wants that phone-channel realism in the data.

      The file is decoded and measured in this browser. Nothing is uploaded or stored.

      Then choose who decides The measurement stays the same. Customer feedback changes who makes the call.

      Measure the phone-hum file to see the check and its route.

      Other files the same checks can test

      These keep the broader bench available without interrupting the main example.

        Make the judgment call

        The number is clear. The meaning isn't. A reviewer listens, decides whether the file belongs in the dataset, and records why.

        Phone-line hum

        Does this file belong in the dataset?

        Measured
        52.2 dB hum contrast
        Automatic check
        Hum is above the 12 dB limit.
        Judgment needed
        Useful phone realism or a broken recording setup?

        See a second judgment call: the natural pause

        Natural pause flagged as a dropout

        The automatic check found silence inside the recording. Listening in context is what tells us whether it is a real transmission gap or someone thinking.

        Automatic route
        returned to contributor
        Status
        open, awaiting review

        See what can ship

        Simulated data

        A review decision changes the shipment, not just the label. The main view shows that consequence. The full batch and formulas stay available underneath.

        See the full simulated batch and the math
        SessionContributorMinutesStateAccepted without a rerecordReview minNote

        Improve the next batch

        Each reviewed decision should leave one useful change behind. That is how the operation gets clearer without quietly changing the customer's standard.

        Natural pause wrongly flagged

        Save the reviewed example

        Future reviewers can hear what this kind of natural pause sounds like before making the same call.

        Owner: senior reviewer

        status: waiting on the natural-pause decision

        Recording level too low

        Add one recording check

        The next contributor brief says to run a ten-second level check before starting.

        Owner: operations

        status: ready for the next contributor briefing

        Phone hum may be realistic

        Take the scope question to the customer

        If this pattern repeats, ask whether phone-channel audio belongs in the dataset before changing the rule.

        Owner: operations, with the customer

        status: waiting on the phone-hum review

        Restore the starting rule, review decisions, consent state, and delivery totals.