AI content pipeline · B2B advisory
Multi-agent AI content pipeline — one case study, three formats, published under human control
DEMONSTRATED LIVE · 29 Jun 2026
Context
A French-speaking advisory firm specialized in company setup in the Middle East wanted a steady, high-end content presence — several LinkedIn posts and Instagram publications a week — without spending hours writing and producing every day. The raw material is rich but sensitive: every case study must be anonymized, hold a senior-consultant tone, and respect regulatory guardrails. The project also served as an AI-mastery showcase: an end-to-end chain, from text to avatar video, governed and verifiable.
Problem
Turning a single case study into three formats by hand — writing, on-brand layout, video script, avatar shoot, subtitling, publishing at the right slot, then moderating comments — is repetitive, expensive work, multiplied across networks. Two non-negotiable requirements: nothing publishes without human approval, and nothing is invented — the system only produces content when it has a real case study in reserve.
Approach
Two tiers of agents. The system itself was built with Claude Code following a disciplined method — brainstorm, design, plan, implementation, tests — then it is operated autonomously by an LLM agent (Hermes, a local Nemotron model) triggered by cron on Mondays, Wednesdays, and Fridays at 9 am: the agent picks the next pending case study, decides which formats to produce, budgets the video script, runs the chain — and stops cold if the backlog is empty rather than inventing a topic. Generation is governed by a single “Brand DNA” — tone, vocabulary, forbidden topics, anonymization rules, and regulatory guardrails — combined with personas injected into every generation, for content genuinely in the brand's voice.
What was delivered
- Two-tier orchestration: an autonomous LLM agent (Hermes) on cron (Mon/Wed/Fri 9 am) that selects the case study, decides the formats, budgets the script, and drives the whole chain — built and tooled with Claude Code (brainstorm → design → plan → implementation → tests).
- Generation governed by a single “Brand DNA” + injected personas: brand tone, anonymization, and regulatory guardrails baked into every prompt, for content in the firm's voice and never naming anyone.
- Complete AI video chain: token-budgeted script → French synthetic voice (ElevenLabs) → realistic 9:16 avatar (HeyGen) → subtitled mp4 assembled with ffmpeg, with frame-by-frame quality control to eliminate the typical defects of AI video (artifacts, lip mismatches, subtitle glitches).
- Three formats from a single case study: a LinkedIn post, a 5-to-7-slide Instagram carousel, and an avatar-video Reel — rendered on-brand.
- Human-in-the-loop: every piece is sent to Telegram as a card (preview + text + ✅ Approve / ✏️ Revise / ❌ Reject buttons); “Revise” actually regenerates the content, it does not merely file the request.
- Publishing and community management: multi-network publishing and database recording through MCP integrations, then an AI community-management agent that replies to comments cautiously and without duplicates (never two replies to the same comment).
- Engineer self-audit: the system is put through a critical audit run by the AI — concurrency, idempotency, security, robustness of LLM outputs — with findings ranked by severity, fixed, then re-validated on an already-proven flow.
- Visual upgrade: a redesign of the editorial charter (before/after) toward a premium identity, applied to every template.
- Source of truth: Supabase (PostgreSQL + Storage + RLS) — case studies, content pieces, publish queue, per-platform posts, and metrics; input form hosted on Vercel.
Stack
- Node.js 22
- Anthropic SDK (Claude Code)
- Hermes agent (local Nemotron model)
- Supabase (PostgreSQL + Storage + RLS)
- Telegram / grammY
- Playwright (visual rendering)
- HeyGen (avatar)
- ElevenLabs (French synthetic voice)
- MCP (multi-network publishing, database)
- ffmpeg (video assembly)
- Vercel (input form)
- crons (Mon/Wed/Fri)
Observed results
On 29 Jun 2026, a single case study was adapted and actually published live — an Instagram carousel and an avatar-video Reel — on a demonstration account, end to end: selection, brand-voice writing, AI video, human approval, then publication.
dated publication · 29 Jun 2026 · demonstration account
Honest about the proof: the demo video is watermarked 720p — it proves the production chain, not the final commercial render. Instagram publication is proven; on LinkedIn, generation is proven and publication remains to be connected.
dated demonstration · 29 Jun 2026
What stays human
Nothing publishes without a human: Telegram approval (✅/✏️/❌) is a mandatory gate, by design. The system never fabricates a topic — with no real case study in reserve, it stops. Client-case anonymization and regulatory guardrails are carried by the Brand DNA and verified as a last line of defense by human review. The firm and the demonstration account are never named.