A fixed pipeline — upload, one transformation, download — is a demo, not infrastructure. Nothing in the portfolio behaved like a general-purpose, extensible media engine: the kind of thing a Cloudinary or a Zapier-for-media actually has to solve. And engineering-first tools like that usually expose their internals — queues, jobs, DAGs — directly in the interface, which reads as powerful to their author and unapproachable to the person dragging a slider.
— Case study 18
Plexus
The edit made by hand on one photo is the exact pipeline that runs across five hundred

project overview
An extensible media-processing engine with a non-destructive, WebGPU photo editor on top, built so the recipe assembled by hand on a single image is the exact pipeline definition a Go worker runs, unmodified, across a whole folder — live preview in the browser, full-resolution execution on the backend, one data structure behind both.
shipped outcome
A working system where dropping a photo into the editor and dragging Light, Color and Sharpen feels instant with no server round-trip, and 'apply to batch' runs that exact recipe as a real DAG pipeline across a whole folder — with live per-step progress and no worker able to silently lose a job. The gRPC plugin contract, which lets a third party add a processor in any language without touching the core, is deliberately phased last because it is additive to an already-working engine, not a dependency of it.
- Role
- Solo architecture and full-stack development
- Timeline
- 2026 — feature-complete case study
- Year
- 2026
- Team
- Solo
Plexus is built around one idea: an edit made by hand on a single photo and a batch pipeline definition are the same data structure. Every editor control writes to a small ordered list of processor-and-parameter steps — a recipe — instead of touching a pixel. That recipe previews live in the browser over WebGPU (with a WebGL2 fallback), and the identical recipe is what a Go worker executes, unmodified, whether it is one image or five hundred. NestJS owns the pipeline DAG resolution and the job state machine; Go owns the CPU-bound ffmpeg and libvips work; NATS JetStream is the one piece of infrastructure behind both job dispatch and the real-time progress stream the front-end subscribes to.
A working system where dropping a photo into the editor and dragging Light, Color and Sharpen feels instant with no server round-trip, and 'apply to batch' runs that exact recipe as a real DAG pipeline across a whole folder — with live per-step progress and no worker able to silently lose a job. The gRPC plugin contract, which lets a third party add a processor in any language without touching the core, is deliberately phased last because it is additive to an already-working engine, not a dependency of it.
— Interface evidence
The product, not a placeholder
These are the verified interface captures from the project build. Scroll inside taller frames to inspect each complete page.








— Delivery process
From operating constraint to shipped system
The work is sequenced around risk. Domain rules and failure states come before interface polish; automation arrives before handoff.
Core pipeline engine
The orchestrator, one Go worker type, Postgres and NATS — linear pipelines only, with built-in resize, convert and compress processors for image, video and audio. No editor yet; the point was proving a job could be dispatched, run and tracked reliably.
Editor MVP
The non-destructive recipe model, the WebGPU/WebGL2 live-preview renderer, and the curated composite sliders — Light, Color, B&W, Sharpen, Crop — with synchronous export. Shippable as its own milestone, independent of the batch machinery.
Real DAGs, realtime and Apply to Batch
Presigned MinIO upload, the shared recipe package, an SSE progress stream off the same NATS events, and branching, parallel DAG resolution — the point where 'edit once, apply to five hundred files' became a real button instead of an architectural claim.
Plugin system — not started
The gRPC plugin contract that lets a third party add a new processor in any language without touching the core. Phased last on purpose: it is additive to an already-working engine, not something the engine depends on.
— System architecture
Clear boundaries, explicit responsibilities
Each layer has one job and a narrow contract. That keeps external services replaceable and product behavior testable.
recipe / one shape
editor = pipeline
one photo by hand = 500 in batch
steps
[ { processor, params } ]
engine / guarantees
Next.js + WebGPU/WebGL2 editor
Upload, dashboard, live progress and the in-browser non-destructive preview renderer — no pixel ever leaves the browser during editing, only the recipe.
NestJS orchestrator
Auth, pipeline DAG resolution and the job state machine: queued, running, partial, complete or failed.
NATS JetStream
A persistent, replayable event log backing both job dispatch to workers and the SSE progress stream to the browser — one piece of infrastructure doing two jobs.
Go worker pool
ffmpeg and libvips-backed built-in processors, horizontally scalable, executing the same recipe at full resolution for export and batch runs.
PostgreSQL + MinIO
Drizzle-managed jobs, pipelines, recipes and plugin registry in Postgres; presigned-URL upload and download through MinIO so the API never proxies large files.
— What shipped
Product capabilities
Non-destructive recipe editing
Every control — crop, light, color, sharpen — writes recipe parameters, never pixels. Undo and redo are just recipe history.
Live WebGPU preview
Slider drags render in the browser with no server round-trip, falling back to WebGL2 where WebGPU is unavailable.
Apply to Batch
The recipe built by hand on one photo runs, unmodified, as a pipeline across an entire folder of files.
Real-time job progress
Per-step progress over SSE, driven off the same NATS event stream used for dispatch, not a separate polling loop.
— Engineering pressure
Challenges resolved
Preview-to-export fidelity drift
The browser's WebGPU approximation and the Go server's ground-truth render must agree closely enough that export never surprises. Drift per composite control is measured numerically against a bound and enforced by a test, not checked by eye.
Crash-safe job dispatch
A worker killed mid-job must not silently drop it. JetStream's persistence, plus explicit ack and heartbeat handling, means another replica picks the message back up instead of the job vanishing.
A language boundary that stays a contract
With Go and TypeScript on either side of gRPC, NATS subjects and a shared Postgres schema, the boundary is easy to let dissolve into ad-hoc JSON. The recipe schema has exactly one canonical definition, and everything else imports or generates from it.
verified project destinations
next transmission
Need a system this deliberate?
Bring the operating problem. We will shape it into a clear product brief and a buildable next step.
Start a project