• Web

FAME Gen Studio

My role: Studio workflow architecture, Angular 20 frontend and NestJS backend development

FAME Gen Studio is an AI content generation and publishing app for brands and agencies. A workspace starts with an AI market research document and a brand kit; from there the user asks the studio in chat for a video, image or text project. An agent-driven pipeline writes the script, plans the scenes, draws keyframes, generates video with the chosen model, adds narration and music, has the result reviewed and stops for approval before it publishes to social platforms. Schedules then repeat the cycle on their own: they generate new projects on a cron, publish them once approved and analyse how earlier posts performed. I worked on how this flow fits together — workspaces, the chat-driven pipeline and its checkpoints — built the Angular 20 frontend that drives it, and worked directly on the NestJS backend behind it.

  • Angular 20
  • TypeScript
  • Signals
  • RxJS
  • Socket.IO
  • Remotion
  • NestJS
  • Tailwind CSS
  • Transloco

Highlights

  • AI pipeline15 typed stages, from research and storyboard to publish and post-publish analysis
  • AutomationCron schedules generate new content and analyse social performance on their own
  • AI models4 video providers, 7 models (Veo, Kling, Seedance) · GPT Image and Gemini for images
  • ProductLive web app for brands and agencies, built at Fame Yazılım
  • Real-time50 typed socket message types · SSE-streamed research documents
  • Frontend scale211 standalone components, 48 shared UI components, 46 spec files

AI Generation Flow

Every request goes through chat. The frontend sends one message over Socket.IO and a backend orchestrator decides which agents and tools to run, so new capabilities arrive without new screens. The protocol types 15 pipeline stages: deep search, scene planning, storyboard and storyboard edits, image generation, video generation, narration, music, scene and film edits, review, publish, post-publish analysis, profile re-pin and analytics.

The run is visible as it happens. Fifty typed message types stream the model's reasoning, multi-step plans (created, step started, step completed, step failed, revised), tool progress from 0 to 100 with asset previews, and live script, storyboard and timeline updates into the editors. A review stage scores the output by category and lists issues with suggested fixes. The user can interrupt a run at any point or limit the AI's edits to chosen scenes.

Human in the Loop

Automation stops where a person should decide. The pipeline raises checkpoints that render inline in the chat and are answered with approve, reject or modify: keyframe approval scene by scene (with a jump into the image editor), publish approval, re-pinning the generation profile and post-publish analysis.

Publish approval shows the AI-drafted caption and hashtags, every target platform with a per-platform format check of the video, and a review warning. Only an explicit approve publishes; a checkpoint that expires is rejected, never published by default.

Research, Brand and Cast

Content is grounded in the brand before anything is generated. Research documents can be re-run on six focus areas at basic or advanced depth, refined with AI over a streamed connection, and exported as markdown. A nine-step branding pipeline goes from research through logo, colours, typography, voice and tone, imagery and motion to guidelines and a brandbook, exported as markdown, Word, poster or slides.

Cast characters keep people consistent across scenes: six face angles and seven emotion references with a completeness score, voice clips in five languages, and positive and negative prompts. Runs that use a cast are routed to GPT Image because it preserves faces. The chosen providers and models are frozen into the plan before a run starts; if the model catalogue cannot be loaded, generation is blocked instead of guessing.

Scheduled Generation and Publishing

Two kinds of schedule run without anyone in the chat. Project-generation schedules hold a prompt, output type, video or image settings, cast and a dated project-name template, and create a new project on a cron in the workspace's time zone, at most once a day. Social-analysis schedules look back over recent posts, compare them with a baseline and save the findings to the workspace knowledge base, at most once an hour.

Each run is logged with its status and, when skipped, the reason (for example not enough credits or a busy workspace), and the user is notified by email or in the app. Before a schedule is saved, a per-run cost estimate comes from the backend; when it cannot be fetched the UI shows no number rather than a guess. Schedules publish on hold by default: auto-publish needs an explicit confirmation, and the agent tool that can create schedules cannot switch it on.

A release calendar tracks every post from draft to scheduled, publishing and published, partially published or failed, polling each platform with back-off. Instagram accounts connect over OAuth today; TikTok, YouTube, Facebook and X are modelled in the release types with their caption and hashtag limits.

Models and Cost Control

Video runs on four providers and seven models — Veo 3.1, 3.1 Fast and 3.0, Kling 3.0 Pro, Seedance 2.0 and 1.0 Pro — and images on GPT Image or Gemini. Each model's limits (durations, aspect ratios, native audio, start/end-frame or reference mode) come from a catalogue in the database, not from constants in the UI. Basic mode is autonomous and leaves the model choice to the backend; Advanced mode pins provider and model at kickoff and exposes six camera, six shot and six lens presets per scene.

Usage is metered in points, where one point is one cent of real provider cost. The monthly plan allowance is spent before a prepaid wallet that carries over, and generation stops when both are empty.

Frontend Architecture

No NgModules: every component is standalone and OnPush, split into core (auth, state, interceptors, tab system, i18n), features (20 feature folders such as editor, cast, branding, schedules, releases), shared (a reusable UI kit of buttons, inputs, modals, drawers, pickers) and a studio layer holding the API service and typed models.

State is signal-first: two global signal-based singletons (app state and auth) plus per-feature stores built from signal() and computed(); a chat state service turns the socket stream into signals, so the UI re-renders only where something changed. An interceptor chain adds the base URL and bearer token and de-duplicates silent token refresh on 401. UI copy runs through Transloco in two languages, with a script that checks key parity.

Editors

A video project moves through three linked phases — script, storyboard, video — in a multi-tab workspace driven by a signal-based tab service instead of router child routes. Tabs toggle with [hidden], so chat history and undo state survive switching between them.

The timeline editor previews a multi-track cut (video, narration, music, overlay) on a Remotion player embedded through a React bridge; cuts, trims, added audio and volume changes are tagged as made by the user or the AI, can be undone and are re-rendered on the server. The image editor runs on a fabric canvas with mask inpainting, and keyframe edits take a separate path so they never cancel a pending approval.

Skills Demonstrated

Skill AreaEvidence in Project
AI product architectureChat-driven agent pipeline, 15 typed stages, inline approval checkpoints, multi-provider model contract
Workflow automationCron schedules, run history with skip reasons, cost estimates, hold-by-default publishing
Angular 20Standalone components, OnPush everywhere, strict templates, lazy-loaded features
Signals & RxJSsignal / computed / effect state, interceptor chain with shared refresh
Real-time UISocket.IO chat with 50 typed message types, SSE streaming with abort
Rich editorsScript, storyboard, Remotion timeline, fabric canvas image editor with inpainting
Design system & i18n48 shared components, Tailwind 4, Transloco TR/EN with parity check
NestJSDirect backend development alongside the frontend features

Want to discuss this project or work together?

Get in touch

© 2026 Abdullah Doğanay