AI-native product company · Auckland, NZ

AI that acts before you ask.

PNR Creations builds CommCopilot, an AI chief of staff for engineering teams. It attends your team's Slack conversations for you, investigates what matters against your codebase, and interrupts you only when your attention or decision is actually required.

See how it worksRequest early access
Slack-first, always onInvestigates your repo, not just the threadEvidence verified before you see it
CCCommCopilot·2 min ago
Action
Why you're notified
A merged change touches code your open PR depends on.
What changed
session/refresh.ts — token refresh now returns a Result instead of throwing. commit a3f9c12 · verified
Impact on your branch
2 call sites use the old signature. evidence: auth/login.ts:48, api/client.ts:112
Likely fix
Wrap both callers in match(); ~15 lines.
Suggested reply
“Thanks for the heads-up — I'll update the two call sites in my PR this afternoon.”
👍👎feedback trains the triage
evidence verified
Product

Stop babysitting Slack. Get decision-ready packages instead.

Built for software engineers whose teams coordinate in Slack — where 90% of channel traffic is noise, and the 10% that matters is only understandable with repository context.

CommCopilot

In development

An AI chief of staff for engineering communication

It attends your work conversations for you, investigates what matters, and interrupts you only when needed.

CommCopilot watches Slack channels and threads, triages every conversation, and — when a message implies your code might be affected — dispatches a coding agent against a read-only checkout of your repository. You get a decision-ready package instead of a ping: what changed, whether you're actually affected, the likely fix, and a draft reply in your voice.

  • Always-on triage: IGNORE · FYI · CONTEXT · ACTION · REPLY · URGENT
  • Autonomous repository investigation by a sandboxed coding agent
  • Evidence validated against git — unverified claims are labelled, never trusted
  • Interrupt budget with a twice-daily digest for everything else
  • Suggested replies drafted in your own writing style
Request early access

How it works

  1. 1

    Listen

    Watches the channels, DMs and threads you scope it to. No prompting.

  2. 2

    Understand

    Classifies every conversation window with a cheap model; deterministic rules score it.

  3. 3

    Investigate

    Dispatches a sandboxed coding agent against a read-only checkout when your code might be affected.

  4. 4

    Decide

    Applies an interrupt budget: silent, digest, notify or urgent.

  5. 5

    Notify

    Delivers a decision-ready package with verified evidence and a draft reply.

Under the hood

A multi-model pipeline via OpenRouter — a cheap classifier tier, a frontier reasoning tier running a bounded tool loop, and headless coding agents. Replay-based evals on labelled conversations gate every prompt and model change.

CC
How we build

AI-native from the first line, not bolted on.

CommCopilot is designed around model inference — and around the engineering discipline that makes inference trustworthy enough to act on.

01 —

Inference before interface

We infer intent from attention and context so the interface can shrink. The best prompt is the one you never had to write.

02 —

Deterministic where possible, LLM only where necessary

Rules drop the obvious cases, models emit structured flags, and code computes the score. Cheaper, faster, and debuggable.

03 —

Evidence is validated, not trusted

Every claim an agent makes about your code is checked against the repository. Unverifiable claims are labelled — never shown as fact.

04 —

Progressive autonomy

Read → investigate → draft → act. Each level of autonomy earns the next through measured accept, edit and undo rates.

Our AI stack

How CommCopilot is built. Provider-agnostic by design.

Multi-model routing

Inference runs through OpenRouter so each pipeline slot — cheap classifier, frontier reasoner — uses the model that wins a bake-off on our own labelled data, not a benchmark.

Bounded agent loops

Tool-using reasoning agents with hard round limits and typed tool contracts: retrieve notes, fetch threads, run an investigation.

Sandboxed coding agents

Headless coding agents (Claude Code, Codex) run against read-only worktrees with stripped environments, timeouts and a fenced JSON output contract.

Replay-based evals

The production pipeline is replayed over hand-labelled datasets. Recall gates block any prompt or model change that regresses.

Company

A small product company in Auckland, New Zealand.

PNR Creations Limited is a New Zealand-registered company that designs, builds and operates its own AI-native software. Our focus is CommCopilot, and we're opening early access to a small group of engineering teams as design partners.

We run lean on purpose: a pure pipeline core that can be replayed and measured, boring infrastructure, and a dependency list we can count on one hand.

Based in
Auckland, New Zealand
Entity
PNR Creations Limited, NZ-registered
Model
Founder-led, engineering-first
Focus
CommCopilot, our AI-native product

Team

PNR Creations on LinkedIn
Get in touch

Want early access?

We're onboarding a small number of engineering teams as CommCopilot design partners. Tell us about your team, your Slack setup and how you work, and we'll be in touch.

Request early access

Contact

Emailphil@pnrcreations.com
LinkedInPNR Creations company page
LocationAuckland, New Zealand