AI Content Factory · Ken Research

LinkedIn Automation
Built Pipeline-First

Research. Approve. Publish. 50 minutes a week.

A production-grade AI system that turns raw market signals into Economic Times-style insight posts across three LinkedIn pages — with you in control at exactly two points in the process.

50m
Human time / week
12+
Posts / month
2
Human gates
Walk through the UI → Architecture

01 · Overview

What this system
actually does

Not a scheduler. Not a template filler. A genuine AI editorial operation — researching at the level of a senior business journalist, with you controlling only what truly requires human taste.

The system ingests live market signals, clusters them into ranked editorial angles, deep-researches each approved topic, writes three post variants, and publishes to the right page at the optimal time — all without you touching it between your two weekly check-ins.

Claude Sonnet 4.5 MCP Connector LinkedIn API v2 Supabase + pgvector Cloudflare Workers
🔍
Signal Ingestion
RSS feeds, news APIs, earnings, macro data — pulled hourly, clustered automatically by theme and structural implication.
🧠
AI Research Agent
Multi-step research loop — web search, gap analysis, 3–5 data points, contrarian angle extraction, everyday analogy.
✍️
ET-Style Drafting
Three post variants per topic: hook-first, data-heavy, narrative. Same research, different editorial posture — you pick one.
🔌
Claude MCP Connector
All Claude calls route through a single inbuilt MCP server. No API keys exposed, no hardcoded integrations.

02 · The Approach

Built as a foundation,
not a one-off

What we're building for LinkedIn is not a LinkedIn tool. It's a content intelligence platform that happens to publish to LinkedIn first.

The architecture — the MCP server, the research agent, the scoring engine, the draft generator, the approval pipeline — is platform-agnostic by design. Every component was built to be reused, not rebuilt.

When we're ready to expand beyond LinkedIn, the work is not starting over. The content intelligence layer stays entirely intact. The only engineering required is a new publishing adapter — a thin connector that speaks each platform's API. The research, the insights, the quality — all of it carries forward.

This is why the investment here is unusually high-leverage. You're not paying for a LinkedIn automation tool. You're paying for the foundation of Ken Research's entire social content operation.

Phase 1 — Now
LinkedIn · 3 Pages
Full pipeline built. MCP server live. Research agent tuned. Approval workflow running. Content factory operational.
Content engine MCP server LinkedIn adapter
Phase 2 — When ready
Twitter / X · Instagram · YouTube
The content intelligence layer is already there. Add a publishing adapter per platform. No rework of research, scoring, drafting, or approvals.
Content engine ✓ reused MCP server ✓ reused New: platform adapters
Phase 3 — Future
Full Social Content OS
One pipeline, one approval workflow, one analytics view — across every channel Ken Research operates. Same 50 minutes of human time. 3× the output.
What is built once and reused forever
What changes per platform (minimal)
✓Signal ingestion & clustering engine
✓T·D·R·N topic scoring model
✓Research agent (web search + synthesis)
✓Draft generation + variant engine
✓Human approval pipeline (Gate 1 + Gate 2)
✓MCP server + Claude connector
↻Platform publishing adapter (API connector)
↻Format rules (character limits, image specs)
↻Platform-specific draft prompt tuning
↻Scheduling window calibration per channel
↻Analytics model per platform API
↻OAuth credentials per channel

"The infrastructure investment is made once. Every platform added after LinkedIn costs a fraction of this build — because the hard work is already done. This is not a tool; it's an operating system for content."

03 · Design Principles

Built on six clear principles

Every architectural and UX decision traces back to one of these. Nothing is arbitrary.

1
⚡
Pipeline is the product
The primary interface is a pipeline view — not a chat window. You navigate by stage, not by conversation. Chat is a support tool within each content item.
2
🎯
Insight over summary
Every prompt is engineered to find the structural implication, the non-obvious angle, and the human-scale analogy. Not a news aggregator.
3
🔒
MCP-first architecture
Claude connects to all external services through a single MCP server. Tool registry is extensible without touching any prompts.
4
📋
Kanban mirrors, not controls
The board reflects pipeline state. State changes happen inside content items — preventing human error from breaking automated sequences.
5
↩️
Fast rejection loops
Rejected drafts re-synthesise from existing research — not from scratch. Revise the essay, not the homework. Under 10 minutes.
6
📊
Scored before you see it
Topics are ranked on T·D·R·N before Gate 1. You see the best angles first, not a flat list. 5 minutes to approve 3–5 topics.

04 · Workflow Pipeline

Seven stages, two decisions

Amber = human gate. Everything else runs automatically. The system never waits on you except at these two precisely designed moments.

01
Auto
Signal Ingestion
RSS feeds, news APIs, earnings releases, RBI/SEBI data, macro indicators — pulled and queued hourly
Hourly · Automated
02
Auto
Topic Shortlisting
Claude clusters signals, identifies structural implications, scores 8–12 candidate angles on T·D·R·N — ranked for review
Daily · Automated
HG1
Human
Gate 1 — Topic Approval
You open the Shortlist screen, review ranked topics with scores and suggested hooks, tap to approve 3–5 topics for the week
~5 min · Weekly
03
Auto
Research Agent
Web search → gap analysis → 3–5 data points → implication mapping → everyday anchor — per approved topic
Per topic · Automated
04
Auto
Insight Synthesis & Draft
Three variants generated — hook-first, data-heavy, narrative — with hook line, citations, page assignment, hashtags
Per topic · Automated
HG2
Human
Gate 2 — Draft Review
You open the Drafts screen, read 3 variants inline, pick one, make light edits in the built-in chat, tap Approve
~15 min · Per post
05
Auto
Schedule & Publish
Posts queued for 3 LinkedIn pages at optimal engagement windows. Engagement data surfaced 48h post-publish
Auto-scheduled · Automated

05 · UI / UX — Full User Journey

Every screen, in order

Below is the complete user journey — from opening the app in the morning to a post being scheduled. Each screen is shown exactly as the user encounters it. Chat is available as a support panel within each content item, not as the primary interface.

Screen 1 ▸ Dashboard — The first thing you see every morning
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Dashboard
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
14
🎯
Shortlist
HG1
🔬
Research
✍️
Drafts
2
📅
Scheduled
✅
Published
System Status
All pipelines running
Good morning — here's today's pipeline
Wednesday, 2 April 2026
⚡
2 drafts awaiting your review
IT Layoffs Pattern · RBI Rate Hold — both scored Strong
📊
Today's shortlist is ready — 10 topics ranked
Gate 1 approval needed for next batch. Top pick: MSME Credit Gap (18/20)
2
Need Review
3
Scheduled This Week
12
Published This Month
4.1k
Avg Impressions
PIPELINE STATUS
📡
Signals
14 queued
🎯
Shortlist
Needs approval
🔬
Research
2 running
✍️
Drafts
2 awaiting you
📅
Scheduled
3 posts
✅
Published
12 live
What you see: Dashboard is your morning briefing. Two attention banners — one for draft reviews (the most urgent human action), one for shortlist approval (weekly Gate 1 task). Stats strip and pipeline status below so you always know where work is sitting without opening individual screens.
Screen 2▸Signals Feed — Raw inputs, auto-clustered
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Signals
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
14
🎯
Shortlist
HG1
🔬
Research
✍️
Drafts
2
📅
Scheduled
✅
Published
Last pulled 18 min ago
Signals Feed
14 signals · 3 clusters identified · Last updated 18 min ago
Cluster A — IT Sector Stress
📰
Infosys, Wipro cut 5,000+ jobs in Q3 — bench utilisation at 5-yr low
Economic Times · 2hr ago · Clustered with 2 others
Strong signal
📈
IT sector billing rates flat YoY despite demand recovery — analyst report
Mint · 4hr ago · Clustered
Supporting
📊
India IT exports growth decelerates to 3.2% in Q3 vs 8.4% prior year
NASSCOM · 5hr ago · Clustered
Supporting
Cluster B — Consumer Slowdown
📰
D-Mart same-store sales growth slips to 5.2%, lowest in 6 quarters
Business Standard · 1hr ago · Clustered with 1 other
Strong signal
🏦
RBI keeps repo rate at 6.5% for 7th consecutive meeting
RBI.org.in · 3hr ago · Clustered
Supporting
Unclustered — Awaiting Shortlister
📡
SEBI proposes new regulations for algorithmic trading by retail investors
SEBI · 6hr ago · Standalone
Pending cluster
📡
India's startup funding hits 3-year low in January 2025
Inc42 · 7hr ago · Standalone
Pending cluster
What you see: Signals are automatically clustered by structural theme before you ever look at them. You can see the raw inputs but you don't need to act here — this screen is informational. The "Run Shortlister" button triggers Gate 1 content generation if you want to manually trigger it before the daily schedule.
Screen 3 · GATE 1▸Topic Approval — ~5 min, once a week. This is your first human decision.
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Shortlist · Gate 1
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
🎯
Shortlist
HG1
🔬
Research
✍️
Drafts
2
📅
Scheduled
✅
Published
Gate 1 open
Topic Shortlist — Gate 1
10 topics ranked · Select 3–5 to approve for research
2 / 5 approved
T · Timeliness
How fresh right now?
D · Depth
3+ data points?
R · Resonance
Senior professional stakes?
N · Novelty
Underreported angle?
01
IT Layoffs — margin play, not demand signal
India's top IT firms are cutting to protect margins amid stagnant billing rates — not because demand is falling
Hook: "India's IT sector didn't fire 5,000 people because demand fell. It fired them because bench utilisation hit a 5-year low."
18/20
Strong
✓ Approved
02
RBI 7th pause — calculated growth bet, not indecision
Seven rate holds in a row is a deliberate growth-over-inflation trade-off that most coverage frames as caution
Hook: "Seven consecutive pauses isn't hesitation. It's a calculated bet that India's growth needs support more than inflation needs fighting."
17/20
Strong
✓ Approved
03
MSME Credit Gap — ₹20Tr structural hole hiding in plain sight
Formal credit to MSMEs keeps shrinking even as the sector grows — supply chain resilience implications ignored
15/20
Good
Tap to approve
04
D-Mart SSS at 5.2% — what consumer wallets are actually saying
Same-store sales slowdown reveals a consumption squeeze that earnings calls are underplaying
14/20
Good
Tap to approve
05
SEBI retail algo rules — protection or market exclusion?
Proposed regulations could price retail traders out of systematic strategies just as they were gaining ground
13/20
Moderate
Tap to approve
Gate 1 — your first decision (~5 min): Topics arrive pre-ranked by score. You read the title, the one-line angle, and the suggested hook. Tap to approve 3–5. Hit "Send to Research" — done. The research agent immediately starts working on approved topics in parallel. You'll see drafts in 30–60 minutes depending on topic complexity.
Screen 4▸Research View — See what the AI found (read-only, fully automated)
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Research · IT Layoffs
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
🎯
Shortlist
🔬
Research
✍️
Drafts
2
📅
Scheduled
✅
Published
Research complete
IT Layoffs — margin play, not demand signal
Research complete · 6 sources · 4 data points · Drafting...
Key Data Points
4 extracted
5-yr
Bench utilisation at lowest in 5 years
Infosys Q3 bench rate 18.4% vs 12.1% in Q3 FY22 — source: Infosys investor presentation
0%
Billing rate growth YoY — flat across tier-1
Avg offshore billing rate $28.4/hr vs $28.1/hr prior year — source: NASSCOM data
5K+
Headcount cuts across Infosys + Wipro in Q3
Infosys 3,400 · Wipro 1,800 · stated reason: "workforce optimisation"
2022
The hiring binge that created today's problem
Both firms hired 40%+ above project demand in FY22-23 anticipating a demand surge that didn't materialise at pace
Structural Implication
This isn't a demand story — it's a margin arithmetic story. When revenue per employee stagnates and hiring costs are sunk, the only remaining lever is headcount normalisation. The next 2 quarters will show whether this is a one-time correction or a structural shift toward leaner IT operations with higher automation penetration.
Everyday Anchor
Think of it as a restaurant that over-hired prep cooks expecting a full house that never arrived. Now they're cutting cooks — not because fewer people are dining out, but because the kitchen ratio got out of hand.
Research Assistant
Ask anything about this topic's research
AI
Research complete for IT Layoffs. Found a stronger margin story than the headline suggests — billing rates are the key variable most coverage is missing.
SB
Any data on attrition rates to compare?
AI
Yes — voluntary attrition actually fell to 12.3% from 23.8% peak in FY23. Counter-intuitive: people stopped leaving right as firms started cutting. Adds to the margin story — costs were stickier than the industry expected.
Drafts generating now...
What you see: Research is fully automated — you don't trigger it. This screen lets you inspect what the AI found before reading drafts. The chat panel on the right is available here as a research assistant — you can ask questions about sources, request additional data points, or probe the angle before reviewing drafts. This is chat-as-support, not chat-as-interface.
Screen 5 · GATE 2▸Draft Review — ~15 min per post. This is your second and final human decision.
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Drafts · Gate 2
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
🎯
Shortlist
🔬
Research
✍️
Drafts
2
📅
Scheduled
✅
Published
Gate 2
2 awaiting review
IT Layoffs — margin play, not demand signal
3 variants ready · KR Insights · Score 18/20 · 1 of 2 drafts awaiting review
Strong 18/20
Hook-first 847 chars KR Insights #ITIndustry #India #LayoffReality
India's IT sector didn't fire 5,000 people because demand fell. It fired them because bench utilisation hit a 5-year low while billing rates stayed flat.

That's a different problem — and it points to a structural shift that most coverage is missing entirely.

When a company's revenue per employee stagnates but costs don't, the only lever left is headcount. What we're seeing is the industry catching up with a 3-year hiring binge that outpaced actual project velocity.

Infosys's bench rate is 18.4% — it was 12.1% three years ago. Billing rates: flat. The maths does itself.

The question to ask isn't "is demand recovering?" It's "at what utilisation level does hiring resume?" Watch that number, not the job cut headlines.
Data-heavy910 charsKR Insights
Three numbers explain everything about India's IT headcount correction:

18.4% — Infosys bench utilisation in Q3 FY26 (was 12.1% three years ago)
0% — growth in offshore billing rates year-on-year across tier-1 players
40%+ — excess hiring growth in FY22-23, anticipating a demand surge that didn't arrive

This isn't a demand story. It's a margin arithmetic story. Revenue per employee has stagnated; the cost base hasn't adjusted.

The correction was inevitable. The question for investors: which firms have the cleaner bench ratios heading into FY27?
Narrative778 charsKR Insights
A restaurant hires 40% more prep cooks expecting a rush that takes longer than expected to arrive. The kitchen fills up; the tables don't. Eventually you cut cooks — not because fewer people are dining out, but because the ratio got out of hand.

That's India's IT sector right now.

The 5,000 jobs cut at Infosys and Wipro this quarter aren't a demand signal. They're the delayed correction to a 2022 hiring binge that assumed project growth would outpace cost growth indefinitely.

Billing rates are flat. Bench is full. Margins were going to get squeezed eventually. They're just getting squeezed now.
Edit Assistant
Request revisions or ask questions
AI
Three variants ready. I'd recommend Variant A — the hook lands harder and the billing rate angle is the most underreported. Variant B works well if you want to attract CFO/investor engagement.
SB
Can you make Variant A slightly more specific about the attrition angle?
AI
Done — added: "Voluntary attrition fell to 12.3% at the same time. People stopped leaving right as firms started cutting — costs were stickier than planned." Variant A updated above.
Gate 2 — your final decision (~15 min per post): Three variants side by side in tabs. Read each, pick the one that fits the page's voice, optionally use the chat panel to request specific word changes, then tap Approve. The "Request Revision" flow sends a note to the AI which re-synthesises from the same research — a new draft appears in under 2 minutes without re-running research.
Screen 6▸Scheduled — Approved posts queued across all three pages
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Scheduled
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
🎯
Shortlist
🔬
Research
✍️
Drafts
1
📅
Scheduled
✅
Published
3 posts queued
Scheduled Posts
5 posts this week · across 3 pages · next publish in 14 hours
KR Insights
2 posts
Tomorrow · 8:30 AM
IT Layoffs — margin play, not demand signal
Variant A · Hook-first
Thursday · 9:00 AM
MSME Credit Gap — the ₹20Tr hole hiding in plain sight
Researching → Drafts pending
Friday · —
Open slot
KR Markets
2 posts
Today · 6:00 PM
RBI 7th pause — growth bet, not indecision
Variant B · Data-heavy
Wednesday · 8:30 AM
D-Mart SSS at 5.2% — consumer wallet signal
Awaiting Gate 2 approval
Friday · —
Open slot
KR Strategy
1 post
Wednesday · 9:00 AM
PE Funds reduce India exposure — what it signals for domestic M&A
Variant A · Hook-first
Thursday · —
Open slot
Friday · —
Open slot
What you see: Three columns — one per LinkedIn page. Each slot shows the scheduled time, post title, and variant used. Slots in amber are approved and queued. Slots with "Awaiting Gate 2" are drafts still pending your review. Empty slots are available for the next batch once Gate 1 runs again. Timing is auto-calculated based on each page's historical engagement windows.
Screen 7▸Published — Live posts with engagement data pulled 48h after publish
👆 On mobile — sidebar & chat panel hidden for clarity. Full layout visible on desktop.
app.kenresearch.ai — Published
Ken Research
Content Factory
Pipeline
🏠
Dashboard
📡
Signals
🎯
Shortlist
🔬
Research
✍️
Drafts
1
📅
Scheduled
✅
Published
12 live this month
Published Posts
12 posts this month · 4.1k avg impressions · Top: KR Insights
4,100
Avg Impressions
3.8%
Avg Engagement Rate
12
Posts Published
47
Avg Human Minutes
PE Funds reduce India exposure — what it signals for domestic M&A
KR Strategy · Published Mon 9:00 AM · Variant A
5,240 impr.
4.2% eng. · 38 comments
India CAD narrows to 1.2% — the story behind the statistic
KR Insights · Published Fri 8:30 AM · Variant B
4,820 impr.
3.9% eng. · 29 comments
Reliance retail underperformance — reading the sector signal
KR Markets · Published Thu 9:00 AM · Variant A
3,860 impr.
3.6% eng. · 21 comments
Manufacturing PMI slide — early cycle signal or blip?
KR Markets · Published Tue 8:30 AM · Variant C
2,410 impr.
2.8% eng. · 11 comments
Narrative variant underperformed
What you see: Published posts with 48-hour engagement data. Impressions, engagement rate, and comment count per post. The last row (amber border) flags a post where the narrative variant underperformed — this feedback automatically weights the scoring model to recommend hook-first or data-heavy variants more heavily for this page's audience going forward.

06 · Architecture

Powered by Claude MCP

All AI calls route through an inbuilt MCP server — a unified tool registry connecting Claude to every service without exposing credentials or hardcoding integrations.

Data Sources
📡
RSS / News APIs
ET, Mint, Reuters, Inc42
📈
Market Data
BSE, NSE, MOSPI, RBI
🔍
Web Search
Live research, live fact-check
💼
LinkedIn API v2
Publish + analytics · 3 pages
↓ HTTPS / SSE
⚙️
MCP Server
Cloudflare Workers · Tool Registry · Auth · Rate Limiting · State
↕ Anthropic API
🤖
Claude Sonnet 4.5
Shortlister · Research Agent · Synthesis · Draft Generator · Gate Manager
↓ Supabase Realtime
💾
Supabase
PostgreSQL + pgvector · Full history
🖥️
Pipeline App
Next.js · All 7 screens · Chat panels
📅
Cloudflare Cron
Signal ingestion · Optimal scheduling
Built-in MCP Connector
Claude connects through MCP

The Model Context Protocol server is the secure bridge between Claude and every external tool. No API keys in the frontend. No hardcoded integrations. Add a new data source to the tool registry — Claude can use it immediately without any prompt changes.

web_searchLive research per approved topic
content_store.writeSave all research, drafts, citations
linkedin.postPublish to 3 pages via LinkedIn API v2
chat.await_approvalPause pipeline, surface gate to human
signals.ingestPull from RSS, APIs, macro data
MCP Config Snippet
// Claude's MCP server registration { "mcp_servers": [{ "type": "url", "url": "https://mcp.kenresearch.workers.dev/mcp", "name": "kenresearch-mcp" }], "model": "claude-sonnet-4-5-20251001" } // Pipeline stage invokes Claude with tools { "type": "tool_use", "name": "web_search", "input": { "query": "IT sector billing rates India FY26", "num_results": 8 } }

07 · Tech Stack

Production-grade from day one

Every component chosen for reliability at scale and minimal operational overhead.

AI Layer
Claude Sonnet 4.5
Anthropic API · claude-sonnet-4-5-20251001
All five automated pipeline stages. Accessed exclusively through the MCP server — never directly from the frontend.
Middleware
MCP Server
Cloudflare Workers · TypeScript
Unified tool registry. All Claude tool calls route here. Handles auth, rate limiting, service-level error recovery.
Database
Supabase
PostgreSQL · pgvector · Realtime
All signals, research, drafts, and publish history. pgvector enables semantic search across content history. Realtime powers live pipeline state.
Frontend
Next.js 14
React · Supabase Realtime · Tailwind
Pipeline app with all 7 screens and contextual chat panels. Server components for fast load; client components for real-time state.
Publishing
LinkedIn API v2
OAuth 2.0 · 3 Page Credentials
Isolated credential sets per LinkedIn page. Publishing, scheduling, and 48-hour engagement retrieval fully automated.
Scheduler
Cloudflare Cron
Cron Triggers · Durable Objects
Hourly signal ingestion. Per-page optimal posting window calculation from historical engagement data.

08 · Timeline

Live in four weeks

Clear weekly deliverables. Week 4 is integration and UAT — no surprises at handover.

Week 1
Foundation
MCP server · Cloudflare Workers
Claude connector + tool registry
Supabase schema + pgvector
LinkedIn API (3 pages)
Signal ingestion pipeline
Week 2
Pipeline Logic
Shortlister + T·D·R·N scoring
Research agent (multi-step)
3-variant draft generator
Gate 1 + Gate 2 approval flows
Rejection → re-synthesis loop
Week 3
Interface
7-screen pipeline app (Next.js)
Dashboard + pipeline strip
Draft review with variant tabs
Contextual chat panels
Scheduled + Published views
Week 4
UAT + Deploy
End-to-end pipeline testing
Prompt tuning with real signals
Voice calibration per page
Deployment on KR infra
Team handover + training