Skip to main content
Live Solution AI · RAG Live · Beta

ContentAI

An AI content engine that generates brand-consistent LinkedIn posts, blog drafts, and email sequences from a brief — in your voice, not a generic AI voice.

A Content Engine That Writes in Your Voice

Give ContentAI a topic and a format. It searches your existing content library to find how you've talked about this before, then generates a draft that matches your tone, your examples, and your audience — not a generic advisory register.

It runs in TechTek's own marketing workflow. The posts, blog introductions, and email sequences published under Hammad's name are drafted here first, then reviewed before publishing.

📝
LinkedIn Posts
150–300 word posts with your framework references and voice
📄
Blog Drafts
Introductions and section outlines matched to your content pillars
✉️
Email Sequences
Subject lines and body copy for outreach and nurture flows

Try ContentAI

ContentAI is in live use at TechTek. A gated beta for select clients is in preparation — request access below.

Beta — Client Access Only

ContentAI — Live Demo

Enter a topic, pick a format, and get a draft in under 3 seconds. Brand voice and ICP profiles are pre-loaded from TechTek's content library.

Request Beta Access

Current beta: LinkedIn posts · Blog introductions · Email subject lines

What TechTek Can Build for You

ContentAI is not a demo environment — it is the production system TechTek uses daily. Building a version for a client's product means starting from a proven architecture rather than a blank page.

  • Brand-aware RAG at concurrency — a retrieval pipeline that returns tonally correct content under simultaneous load, not just topically similar chunks.
  • Real-time streaming output — draft content appears word by word in under 3 seconds, not as a delayed bulk response.
  • Quality gating before output — generated drafts pass an automated review step before being shown. The model critiques its own output; if it fails, it retries before the user ever sees a bad result.
  • Feedback loop that improves over time — published content and engagement signals feed back into the retrieval store, so the system learns which outputs perform.

Three Steps, Under 3 Seconds

  1. 1
    You give it a brief — a topic, a format (LinkedIn post, blog intro, email), and the audience you're writing for. That's it.
  2. 2
    It finds your voice — the system searches your existing content library to find the posts, articles, and frameworks you've used before on this topic. It uses those as context, not as copy to repeat.
  3. 3
    It streams a draft — the output appears live, word by word. Before publishing, you review it. If you improve it, that feedback makes future outputs better.

Tech stack

Next.js 14 Supabase pgvector Claude Sonnet Vercel Edge Upstash SSE Streaming

Four Decisions That Determined the Architecture

These are the choices that had real consequences — not the obvious ones.

Decision 01

Two-stage retrieval over simple similarity search

Fetching the top-20 most similar chunks floods the prompt with content that is topically related but tonally wrong — old drafts, off-brand fragments. A cross-encoder re-ranks the candidates by tonal fit before prompt assembly. Result: tonal fidelity improved from 61% to 89% in blind review. Token use dropped 60%.

Decision 02

Claude Sonnet over GPT-4

GPT-4 produced accurate content but defaulted to a generic advisory register — detectable as different from an established author's voice. Claude Sonnet held tonal consistency across long-form posts. Cost was also 4× lower per generation. Switched in Week 2 and not revisited.

Decision 03

Server-Sent Events over WebSockets

WebSockets added persistent connection management and edge-compatibility complexity for a UX outcome identical to SSE streaming. Next.js App Router handles SSE natively. The code is 60% simpler. The user experience is the same.

Decision 04

Deterministic prompt budgets

Unconstrained retrieval causes inconsistent output quality under concurrent load — some users get tight context, others get noisy prompts. The pipeline is designed so every request arrives at the generation step with a prompt between 1,100 and 1,400 tokens, regardless of what retrieval returns. Output quality is predictable at any load.

Build this with TechTek

Building something similar?

Whether it's a content engine, a document processing pipeline, or any system that needs to generate brand-consistent output at scale — this is the architecture we've proven in production.

Let's talk →