Whisper Portal — Transcription & Agentic Code Editor

A bulk transcription platform that tells speakers apart and grew into an AI code editor, where the model is given real search, read and edit tools and investigates a codebase instead of guessing from a pasted file.

Letting an AI investigate the code before it edits

216,604lines of code
334files
615commits
12data models
Screenshot of Whisper Portal — Transcription & Agentic Code Editor

The hard part

Most AI code editing pastes a file into a prompt and hopes the answer is in it. When it is not, the model produces something plausible and wrong, much like answering an exam question without reading it. A cleverer prompt does not fix that. The model needs to be able to look for itself, and to be trusted to decide where to look.

The model is given search, read and edit as real tools and runs the investigation itself: it decides what to examine, fetches it, and then makes the change. Around it sits a transcription platform. Meetings are recorded until someone stops them, speakers are separated using a fleet of machines, and eleven analyses are available on demand. All of it sits behind a login that denies access by default, with a decoy that hands automated scanners convincing fake credentials.

561 commits on 59 working days since 2026-02-28; 8 separate source layers; 6 local inference engines.

What the measurements mean

Only what is unusual for this project. Every sentence names the field it came from; nothing here is asserted that cannot be checked against the data.

How it works

Most AI code editing pastes a file into a prompt and hopes the answer is inside it; when it is not, the model invents something plausible. This one runs a genuine tool-calling loop. The model decides what to look at, finds it by searching the codebase, reads the symbols it needs, and only then makes the edit, so each change comes from an investigation rather than a single guess. Around it sits the original transcription pipeline, with swappable backends for diarization (working out who spoke when). It now covers meetings too. A meeting records until it is stopped, and meetings transcribed on the fleet get speakers by diarizing first and then transcribing each speaker's stretch. A finished meeting offers eleven analyses on demand, and questions about it are answered from the transcript. It is served through a login-gated portal behind a reverse proxy, and access is denied by default, so every route is closed unless it is deliberately opened. A honeypot layer returns convincing decoy credentials to automated scanners instead of a 404. An Exchange offers a download-only view that a browser can open with an API key.

What it is built from

Each bar is one of this project's own modules, sized by how many times the rest of the project imports it. The counts are measured, not estimated, and the module at the top is the one the rest of the project leans on most.

Module mapInternal modules by how often they are imported: llm_handler 90 times, summarizer 27 times, code_retriever 23 times, agent_editor 23 timesllm_handler90summarizer27code_retriever23agent_editor23imports of each internal module, counted across the project

What kind of engineering this is

Six signals counted from the source: files that run models locally, that act without being asked each time, that check their own output, that survive failure, that touch parsers and protocols, and that were measured then improved. Every axis carries its file count.

Depth profileDepth signals: Local inference 42 files, Autonomy 24 files, Verification 54 files, Resilience 137 files, Low level 52 files, Iteration 0 filesLocal inference42 filesAutonomy24 filesVerification54 filesResilience137 filesLow level52 filesIteration0 files

What it is written in

Share of code bytes. Vendored bundles and generated output are set aside before this is computed, which is why it differs from a simple file count.

Language mixPython 84.3%, HTML 14.4%, JavaScript 0.6%, CSS 0.3%, Batchfile 0.3%, SVG 0%Python 84.3%HTML 14.4%

What stands out

Capabilities

Grade

Every check names the field it measured. Checks that cannot apply are excluded, never counted as failures.

B
82.9% over 19 of 19 checks100% coverage

Not met

Architecture

Directories carrying real modules, largest first.

Data model

Typed structures and tables found in the source.

Vulnerable dependencies

20 declared dependencies match a published advisory. Checked offline against the OSV database; only pinned versions can be checked, so anything unpinned here is unknown rather than clean.

Vocabulary

Parsed 268 of 268 Python file(s). Counts are exact; the names below are a sample.

3,720functions
1,982methods
225classes
268documented modules

Classes

Functions

Command line

Depth

Signals counted from the code, not asserted.

615 commits across 67 separate working days over 8.5 months, 204 of them revising existing behaviour

Built from scratch

Closes its own loop

Proves itself

Survives failure

Local inference

Systems

Working within limits

Environment engineering

Distributed systems

GPU & memory management

Model operations

Languages

Built with

Recurring themes in the commit history

Evidence

Taken from the code itself: a module docstring or the opening of the README.

Personal benchmark: mine the owner's OWN commit history for agent tasks. A case is a real commit of one of the owner's repos that changed source code AND a test, where the commit's version of that test FAILS on the parent commit and PASSES on the commit itself - in that repo's own interpreter. The task is the commit's message; the commit's test is the hidden grade. Cases carry the owner's conventi

At a glance

Status
active
Primary language
Python
First commit
2026-02-28
Last activity
2026-11-13
Domain
AI & machine learning
Agent docs
CLAUDE.md, AGENTS.md, PROJECT_KNOWLEDGE.md, .claude, .superpowers