Skip to content
Built at Build Out Loud, Bengaluru

Meeting mein bola.Repo mein bana?

A voice teammate for engineering teams. It remembers what was decided, asks the lead when it isn’t sure, and opens the exact line of code. Every answer comes with its source.

  1. Team MemoryWhat the team decided←
  2. Leader DeskAsk the lead↑
  3. Code NavigatorFind it in the code→
You

“Code freeze kab hai, 3:30 ya 3:45?”

Team Memory

That was only proposed, not confirmed. I won't guess. Want me to ask Charan, the lead?

Proposed, not confirmedWispr notes · Sprint sync, 11:04 · passage 11
←→ switch agents ↑ ask the lead
Runs onWispr FlowBolnaGitHubPick an agent withAirPods head turnWebcam←↑→Keys
Demo path, drawn on this repository’s real code graph
Every answer walks a path you can see.

One question, traced end to end: from voice, to the meeting line that decided it, to the exact function in this repo, to the branch that fixes it.

  1. Voice in
  2. Understood
  3. Wispr MCP → meeting line
  4. Graphify → file → function
  5. Test → branch → PR
Harsh asks out loud. Bolna turns the voice into text.
background: 454 real nodes from harshsinha-12/kaam-ki-baat @ 317a348

Meetings end. Doubt doesn’t.

A decision gets spoken, never written down properly, and an hour later three people remember three different versions. These are the numbers behind that feeling.

68%

of people don’t get enough uninterrupted focus time.

Microsoft Work Trend Index 2023, 31,000 people · source
57%

of the workday goes to meetings, chat and email.

Microsoft Work Trend Index 2023, Microsoft 365 usage · source
55%

say next steps are unclear when a meeting ends.

Microsoft Work Trend Index 2023 · source
61%

of developers spend 30+ minutes a day searching for answers.

Stack Overflow Developer Survey 2024 · source

Every team runs the doubt loop. Nobody owns it.

The decision was spoken, then never written down properly. The code changed, but nobody checked which commit. The one person who knows is busy. So people guess, and guesses ship.

Someone has been waiting 0:00 for an answer.
  1. Meeting ends“Someone will own it.”
  2. Slack ping“Wait, who owns this?”
  3. WaitingThe lead is in 4 meetings
  4. A guess“I think it was merged.”
  5. ReworkWrong commit, old CI

Wispr hears it. The evidence decides it.

Half-sentences and Hinglish are fine. What you say is cleaned up, matched to a decision or a function, checked against fixed rules, and only then answered.

Where you look decides who answers.

Turn left, face front or turn right with AirPods, a webcam or the arrow keys. A turn only picks the agent. You still choose when to talk.

You

“Code freeze kab hai, 3:30 ya 3:45?”

Team Memory

That was only proposed, not confirmed. I won't guess. Want me to ask Charan, the lead?

Proposed, not confirmedWispr notes · Sprint sync, 11:04 · passage 11
←→ switch agents ↑ ask the lead

One desk to check that the whole team got it.

Ask the Leader Desk once. It calls each teammate through Bolna to confirm they understood, and walks the code graph to show what actually changed in the repository since the meeting.

  • Calls, not reminders. Each teammate hears the decision and confirms it in their own words.
  • Follow-ups come back to you. Open questions land in your inbox, never as a guess.
  • The code is the proof. The graph walk ends at the exact lines that changed, with the test that covers them.
Leader Desk
CR

Team checkvia Bolna calls
HSHarshWaiting
NKNehaWaiting
AMAmanWaiting
Code since the meetinggraph walk
DecisionCode freeze 3:45TaskGate saves on CIPR#42 mergedFilepersist-update.tsFunctionpersistVerifiedUpdateTestpersist-update.test.ts
lib/core/persist-update.ts PR #42 · CI passing
35 export async function persistVerifiedUpdate(task: Task) {
36+ const ci = await checks.forCommit(task.sha)
37+ if (!ci.green) return { saved: false, reason: "CI not green" }
38 const saved = await store.write(task)
39 return store.readBack(saved)

From “not sure” to a sourced answer.

Unknowns become a question for the lead instead of a guess. The answer comes back with its source, even if you hung up in the meantime.

Confident is easy. Correct is the product.

Six questions where a general chatbot sounds helpful and is wrong. Each refusal below is a rule in the code, with a test.

You askA typical chatbotKaam Ki Baat
Is it merged?Yes, the PR is done.PR #2 is open, not merged.An open PR is never reported as shipped.
Did CI pass?Yes, all green.Green on a1b2c3d, an older commit.A pass on an older commit does not verify the current one.
Am I the owner?Yes, you were assigned.Only proposed so far. Should I ask Charan?Mentioned, assigned and accepted are kept apart.
Mark it done.Saved! (it had timed out)Saved and read back, version 3.A timeout is a failure, and it says so out loud.
I turned to the leader desk.Welcome, leader.Selected. Confirm with your passkey.A head turn only selects. Identity is a server role plus a passkey.
(cut off mid-sentence)The fact is lost.Also, before I forget: freeze needs green CI.Unsaid critical facts come back after an interruption.

“Kahan hai?” gets a line number.

Code Navigator walks the repository graph to a file, a function and the nearest test, pinned to one commit.

harshsinha-12/kaam-ki-baat · lib/core/persist-update.tsConfirmed at 63317bd
25export async function persistVerifiedUpdate(args: {
26 raw: unknown;
27 trustedActorId: string | null;
28 authorizedTaskIds: ReadonlySet<string>;
29 repository: UpdateRepository;
30}): Promise<PersistResult> {
31 const parsed = parseTaskUpdate(args.raw);
32 if (!parsed.ok) return { saved: false, reason: parsed.reason };
33 const req = parsed.value;
34 if (!args.trustedActorId || !args.authorizedTaskIds.has(req.taskId))
35 return { saved: false, reason: "unauthorized" };
Next: npm test -- tests/persist-update.test.tsOpen at 63317bd ↗

Measured, not guessed.

Numbers from our own repository and our own calls. Read them as a regression guard, not a public benchmark.

Code search

Right function in the top 3 · 28 questions
n/a
grepfiles only
46%
Keyword (BM25)baseline
39%
First graphone seed
93%
Semantic graphours

Held-out set of 12 questions never used for tuning: 33% vs 17%. Full eval

Tests on every push

GitHub Actions, 15 Nov · 0 failures
148 at 10:14223 / 223 passing

Every call, evaluated

7 Bolna attempts · 45 s quality gate
45 s
Phone
48s
Phone
94s
Phone
50s
Browser
158.5s
Browser
102.1s
Browser
11.3s
Browser
0s

The 94 s call ended with a voice-provider error. Each call has a written review. Open call evals

Sources in. Evidence out.

Wispr supplies reviewed passages from one approved meeting, GitHub supplies commit-pinned code. Fixed rules decide status; the language model only explains.

Wispr Flow MCPone approved meetingGitHubone repo, commit-pinnedLive speechopt-in, provisionalHead turn or tapAirPods, webcam, keysEvidence engineroles, consent, reconcileversioned memorywork and call ledgersTeam Memoryleft, membersLeader Deskcentre, lead onlyCode Navigatorright, repo accessBolna voicephone, 5 HTTPS tools

Works now

  • Evidence engine, three agents, leader escalation with read-back
  • Live GitHub reads: PRs, changed files, check runs
  • Bolna runtime tools over HTTPS, tested with fixtures
  • AirPods motion companion and webcam controls with calibration

Next

  • Verify the voice path in a real call
  • Three consented phone calls longer than 45 seconds
  • Passkey checks on real devices
  • Optional realtime sync once tests match

Stop guessing what the team meant.

Open the workspace, connect a repository and ask your first question.

Kaam Ki Baat: Meeting mein bola. Repo mein bana?