Admin guide
PULSE (Prescriptive Unified Lead & Social Engine) by Runink keeps a marketing team’s daily work moving; this guide is for the people who deploy and operate a PULSE instance.
Architecture
| Part | What it is |
|---|---|
| Backend | One Go server (pulse-server serve). gRPC, gRPC-Web for browsers, and a few plain HTTP routes share one port. |
| App | Flutter, an Android and iOS app. The backend serves no web bundle. |
| Data | No database. Every table, account, session and media file lives in PULSE’s sealed appfs root on the cluster’s own object store (objectd, bucket pulse). There is one writer, so there is one backend instance per deployment. The server refuses to start if it cannot mount the root. |
| Text model | The platform’s own inference: PULSE calls the model router (INFERENCE_REMOTE_URL), which sends each request to a tier served by Distro-built mistral.rs on the GX10 inference servers. No third-party model or API is ever called. |
| Product facts | Read straight from the files in grpc/data/corpus by search and file reads (direct corpus interaction). There is no vector index and no similarity search. |
| Speech | Text-to-speech through TIDE’s shared ttsd (TTSD_URL); speech-to-text through the voice tier (VOXTRAL_REMOTE_URL). PULSE embeds no speech engine. |
| Secrets | OAuth tokens and passwords are sealed with the envelope key TIDE_ENVELOPE_KEK before they reach the database. |
There is no image-generation tier: PULSE draws carousel slides and video frames itself in Go and assembles video with ffmpeg.
Who deploys PULSE
Runink TIDE does. This repository holds application code and the backend image build. Cluster bring-up, the operator and deploys live in TIDE:
# in TIDE's repository, from its grpc/ module
go run ./cmd/tide deploy pulseTIDE’s pulse-operator reconciles the instance into the pulse-system
namespace from a PulseControlPlane resource. That resource is the
authoritative pod spec: change it or its overlay, never the live Deployment,
which the operator puts back. TIDE’s k0s overlay sets PULSE_ENV=production.
The names on the cluster:
| Object | Name |
|---|---|
| Deployment (named after the PulseControlPlane) | pulse, so kubectl -n pulse-system rollout status deploy/pulse |
| Pod label | app.kubernetes.io/name=pulse-backend |
| Container | backend |
| Image | pulse-backend, built by this repository’s cd-backend.yml |
There is no deploy/pulse-backend.
How a change reaches production
cd-backend.yml builds the image on every push to main, and again every
night at 05:23 UTC (the shared org-runink/ui components are compiled into
the image, and its path filter cannot see them). It then compares the digest
it pushed with the digest the pod runs and rolls only when they differ (or
when it cannot tell): it deletes the pod by its label and waits on
deploy/pulse. A pod delete changes no pod template, so the operator has
nothing to undo, and imagePullPolicy: Always makes the new pod pull the new
image.
Which commit is running is stamped on the live PulseControlPlane, not on
the image (werf tags it :latest) or the pods (the operator rewrites their
labels):
kubectl -n pulse-system get pulsecontrolplane pulse -o yaml | grep 'runink.org/'runink.org/commit, runink.org/image-digest, and runink.org/rolled-at,
rolled-by, rolled-on. rolled-at is written only on success, so its age
is the staleness signal: more than about 36 hours means the runs are failing
or the schedule stopped. When you compare digests, read the container named
backend.
What this repository ships
The backend image, built from grpc/Dockerfile with the repository root
as the build context. The build is self-contained: it clones the sibling
libraries grpc/go.mod points at at pinned commits, and compiles a static,
cgo-free Go server. There is no Flutter or web stage; PULSE ships as an Android
and iOS app. Every base image is pinned by digest and comes through the
BASE_REGISTRY build argument; see
Image build arguments.
The runtime image is debian:bookworm-slim with ffmpeg (Reels, Stories
and other videos), chromium (everything that reads live web pages: the
audit, lead prospecting, market signals). The entrypoint is pulse-server serve; nothing runs beside it and nothing durable is kept in the pod.
First things to set
| Setting | Why |
|---|---|
PULSE_JWT_SECRET | Required. The server will not start without it. |
TIDE_ENVELOPE_KEK | Required. The database refuses to open without it, rather than write tokens in plaintext. |
TENANT_ID | Names the tenant for billing, the profile page and the Admin page. |
CORS_ALLOWED_ORIGINS | Required in production. The backend serves no web bundle, so without PULSE_WEB_DIR a production deployment must set this to the pinned origins that may call the API (no wildcard) or the backend refuses to boot. Set it in the deployment before this change rolls out. |
INFERENCE_REMOTE_URL, INFERENCE_MODEL | The model router, and the model name requests carry. |
PULSE_PUBLIC_BASE_URL | The https address people use. Needed by every sign-in connection, Instagram media and the podcast feed. See Public address and redirects. |
Then work through Switches and arming order. The full list is in the configuration reference.