01 — 6 capabilities

Multi-agent generation pipelines

Four production pipelines, each a chain of specialist agents on its own queue — so a slow blog run never blocks a LinkedIn post.

LinkedIn post pipeline
Hook Finder → Post Writer → Editor → QA Checker → Approver. Returns the post copy, alternative viral hooks, line-break shaping and a hashtag cluster. Runs on a dedicated queue at concurrency 4.
Technical blog pipeline
Web Researcher → Section Writer → Editor → QA Assembler → Approver, producing a full technical markdown post with working code blocks. Concurrency 2, because research is slow and expensive.
YouTube script pipeline
Topic Researcher → Script Writer → Editor → QA → SEO Specialist. Emits a full video script, a Shorts cut, high-CTR title options, tag sets and thumbnail concepts.
Instagram visual synthesis
SDXL Turbo / Flux.1 Schnell through a local ComfyUI instance, rendering 1080×1350 4:5 post backgrounds on a serialised GPU queue so the card never contends with itself.
Live agent pipeline monitor
Every agent step streams to the dashboard over a WebSocket as it happens — which agent is running, what it produced, how long it took. No polling, no black box.
Multi-provider LLM fallback
Ollama (local), NVIDIA NIM, Groq, OpenAI, Anthropic and Gemini behind one factory. Pick a provider per run, or let the chain fall through when one is down.

02 — 7 capabilities

Content operations

The part most tools skip: what happens between the idea and the upload.

Editorial planner
Weekly and monthly plans with AI-suggested topics, deduplicated against every topic you have already published. Publish dates are assigned deterministically in code — never left to the model.
Six-stage production tracker
Each plan item moves through pending → script generated → presentation created → video created → video edited → video published, colour-coded down the list.
Uncovered topics backlog
Auto-research generates a batch of candidate topics per run. The first is used immediately; the rest are saved as an 'uncovered' backlog instead of being discarded — deduplicated per user, generated on demand later.
Approval queue & content review
Nothing publishes itself. Generated drafts land in a review queue with the full body, and each piece is approved or sent back per platform.
Google Calendar sync
One-click event creation for approved items, with 1-day and 1-hour reminders, re-synced automatically when a plan item's date moves. Tokens are Fernet-encrypted before they touch Postgres.
Planner analytics
Production pipeline by stage, platform mix, review status, content volume over time, schedule adherence, generation reliability and average words per piece — across every plan.
Telemetry ledger
Every pipeline run recorded with its provider, status, duration and per-agent step trace. Failures stay on the record rather than disappearing.

03 — 9 capabilities

Live channel data from the Data API and the Analytics API, then Gemini on top of the real numbers — not on top of a guess.

Live upload sync
Every upload pulled from the YouTube Data API v3 — title, tags, duration, views, likes, comments — into Postgres, plus a nightly Celery Beat refresh at 12:05 AM IST.
Real impressions & CTR
A separate pull from the YouTube Analytics API for impressions, click-through rate and average view percentage — the metrics the public Data API will not give you.
Per-video SEO audit
Title length, tag count and description word count scored per video, so the weak listings surface without scrolling Studio.
AI channel audit
A full Gemini 3.6 Flash performance audit of the channel — what is working, what is not, and why. Persisted, so it does not regenerate on every page load.
Channel strategy brief
What's working, where the gaps are, what to prioritise next — grounded in the synced performance data rather than generic advice.
1–8 week publishing roadmap
A dated, week-numbered publishing plan for a window you choose. Four weeks is the practical '30-day content plan'; items push straight onto the production board.
Per-playlist trend suggestions
'What should I make next for this playlist' — grounded in live web search alongside the playlist's own performance.
Per-video growth recommendations
Title, thumbnail, hook and tag fixes for one specific video, grounded in that video's real stats plus current trends.
Playlist production board
A local idea → published tracker for planned videos, kept separate from real YouTube playlist writes.

04 — 6 capabilities

A full Meta OAuth implementation with a token lifecycle that actually survives, and a three-tier read path that keeps the dashboard fast.

Meta OAuth 2.0 + 60-day tokens
Short-lived code exchanged for a 60-day long-lived token, stored Fernet-encrypted, and auto-refreshed whenever fewer than 15 days remain. Connection health is visible on the page.
Three-tier analytics read path
Redis cache first, a pruned Postgres snapshot second, and the live Meta Graph API only on an explicit sync or the nightly 12:10 AM IST refresh. Rate limits stay comfortable.
Profile, reach & post performance
Followers, total reach, impressions, link clicks, average reel views and carousel saves, with a ranked table of top reels and carousels by engagement rate.
Scored hashtag clusters
Every hashtag you have used, grouped and scored by the average reach of the posts that carried it — paginated across the full set.
AI strategy brief & roadmap
Gemini-generated strategy briefs and 1–8 week content roadmaps, grounded in the synced profile, post and hashtag data, and persisted between sessions.
Instagram content board
A Kanban board across idea → in production → scheduled → published, with image-prompt history, scheduled dates, and a one-click bulk 'add roadmap to board'.

05 — 8 capabilities

The unglamorous half — which is the half that decides whether any of the above survives contact with production.

First-party cookie/JWT auth
HttpOnly access and refresh cookies, CSRF double-submit, a Redis revocation denylist and email OTP verification. No third-party auth provider in the path.
Boot-time auth coverage guard
Every route must declare authentication or be explicitly allowlisted, or the application refuses to start. A missing dependency is easy to miss in review; a boot failure is not.
Split Redis topology
An evictable LRU cache instance for analytics is kept separate from the AOF broker instance holding Celery state and the auth denylist — so a cache eviction can never drop a queued job.
Distributed Celery workers
Dedicated workers per queue — default, LinkedIn, blog and a serialised GPU queue — plus Celery Beat for the nightly YouTube and Instagram syncs.
Vector memory & RAG
Qdrant-backed document ingestion, semantic search and per-user agent memory, so the agents can be primed on your own material.
Async API, sync workers
Two engines against one database: asyncpg for the FastAPI request path, psycopg for Celery. Workers get a real pool instead of an event loop per task.
Versioned schema
25+ tables under linear Alembic migrations, applied by a one-shot migrate service that completes before the API or any worker starts.
Rate limiting & usage tracking
Per-user quotas and slowapi rate limiting in front of the generation endpoints, so one enthusiastic tab cannot drain your provider credits.