Instagram is the storefront. Nothing was built for it.
The brand was managing Instagram across posts, reels, carousels, ads, stories and collabs with a stack of native tools: Meta Business Suite for one thing, a third-party scheduler for another, Google Sheets for SLA tracking, and a Telegram group for crisis comments. Comments were missed. Sentiment was guessed at. There was no source of truth.
One platform, one inbox, one feed.
Meta webhooks push every comment, mention, and DM into a Node ingestion layer the moment it happens. Token threading and automatic media refresh handle Meta's quirks. PostgreSQL stores the canonical event log; SSE pushes updates to the frontend so the inbox feels live, not polled.
- Real-time comment pipeline via Meta webhooks — HMAC-verified, with error recovery and replay. 1,632 comments ingested and scored to date across 234 posts, reels, carousels and ads.
- GPT-4o-mini sentiment scoring per comment: positive, negative, question, neutral, spam, crisis — flowing into trend dashboards with a plain-language mood axis.
- Permanent media cache — Instagram's CDN URLs expire within hours, so every reel and thumbnail is mirrored to disk, with a self-healing worker sweeping for gaps every 30 minutes.
- Eight background workers on cron: sentiment every 2 min, SLA tracking every 5, anomaly detection every 15, media healing every 30, token refresh daily.
- SLA-tracked inbox with unanswered detection, reply workflows, and SSE-powered live updates — no polling, no refresh button.
A side-quest that became a moat.
Building igdesk solo meant constantly hitting context-window walls in AI coding sessions. So I built graphify: a codebase knowledge graph that produces semantic subgraphs of just the relevant code for the task at hand. It gets injected into AI coding sessions — cutting token usage by roughly 75%.
Boring stack, sharp execution.
Stack chosen for reliability and ownership, not novelty. AWS EC2 Tokyo. Caddy for automatic HTTPS. Cloudflare DNS. systemd for service management. Daily Postgres backups.
Every surface handled. Every voice heard.
The social team works inside igdesk now — collabs, reels, ads, mentions and the inbox, one screen. Nothing gets missed: every one of the 1,632 comments that arrived has a sentiment score and an answer state, and a crisis comment surfaces the moment it lands. The listening layer turns comment noise into a mood the brand can act on — what people love, what they question, what's about to become a problem — right where monthly content reviews happen.