~/projects/tala/README.md
TALA is a self-hosted work hub. It replaced obligations scattered across chat, Google Tasks, email and two internal apps with one Postgres-backed source of truth. It flags stale staff tasks, turns spoken notes into reviewed tasks, mirrors my own tasks to Google Tasks, and runs an email loop where staff reply with photo proof before a task is closed.
cat architecture.flow
architecture.flow
Capture
Staff loop
Access
cat problem.md
Work obligations lived in too many places: chat threads, Google Tasks, email, the Liquidation Tracker and Project ATLAS. Nothing showed what I owed, what my staff owed, and what was waiting on someone else.
Delegated tasks had no notion of "last update". A task assigned on Monday could go quiet for a week before anyone noticed, and following up depended on memory.
Capturing work while away from a keyboard (in the field, in a meeting, driving) meant losing it or typing it in later.
cat solution.md
I built TALA as a single-user web app with PostgreSQL as the one source of truth. Every task, follow-up, project and deadline lives there. A task assigned to staff that has had no update for three days is automatically flagged as stale.
Voice capture records up to two minutes in the browser and transcribes it on a local faster-whisper service, so audio never leaves the homelab. A purpose-built n8n workflow then turns the transcript into proposed tasks. It resolves staff names ("ask the technician to…") and leaves any ambiguous name unassigned instead of guessing. Nothing is created until I approve the proposed tasks on a review card.
Staff follow-ups run as an email loop. A weekday-morning job queues one digest per person, n8n sends it, and staff reply with photo proof. A reply can never close a task. Only my explicit Accept does.
A read-only "Pendings" panel shows what is waiting in the Liquidation Tracker and ATLAS, mapped to who it is waiting on, so I don't have to open three apps every morning.
ls ./components
One Next.js 16 app serves the UI and talks to PostgreSQL directly through route handlers. That keeps a single-user tool to one container and one deploy.
Cloudflare Access gates the hostname with Google sign-in. The app's proxy re-verifies the Access JWT's issuer, audience and email, and fails closed if any config is missing.
n8n talks to TALA through an internal API protected by a bearer token compared in constant time. Requests arriving through the public tunnel are rejected outright, so it is reachable only on the Docker network.
A generic "run any prompt" relay was rejected as an arbitrary-execution surface. Each LLM job is its own n8n workflow with a fixed prompt. TALA sends only quoted text and staff names, and re-validates everything that comes back.
A faster-whisper service runs on CPU inside the homelab with a vocabulary prompt for mixed English and Tagalog. It handles one request at a time and returns 429 when busy, so it can't starve other workloads.
Six numbered SQL migrations are tracked in a schema_migrations table. The database runs under its own least-privilege role and is included in the nightly backup. Proof photos are stored in Postgres so they are backed up too.
./run --trace
MediaRecorder captures up to two minutes of audio in the browser.
Audio is posted to the local Whisper service. Nothing goes to a cloud speech API.
The Parse Note workflow runs a fixed prompt through Claude with the current staff list, and returns structured task proposals.
TALA re-validates the JSON and matches names. Ambiguous assignees stay unassigned rather than guessed.
A review card shows every proposed task. Nothing is written until I click "Add N tasks".
./run --trace 2
A weekday 08:00 (Manila) job writes one reminder digest per staff member to a mail outbox table.
n8n polls the outbox every two minutes and sends over SMTP. TALA decides every recipient and body, and n8n only transports.
Staff reply by email with photos. IMAP intake de-duplicates by Message-ID and checks each image by its magic bytes and size limits.
I review the proof and choose Accept & close or Send back. A reply alone never closes a task.
tail -f incidents.log
cat package.json | jq .stack
echo $EXIT_STATUS
outcome.log
TALA is my daily driver. Every task I own and every follow-up I delegate is in one place, stale work surfaces on its own, and spoken notes become reviewed tasks in under a minute.
The project is also a deliberate exercise in safe AI automation. The LLM has a narrow, fixed job; a human approves every write; outbound email is gated behind a preview mode the tests enforce; and every external integration was verified against the real service, not just mocks.