Targeted-Marketing-Using-LLM
Contentify — LLM-Powered Marketing Content Scorer
The Problem
Marketing teams spend hours crafting content — only to find out after launch that it didn’t land with the audience they were targeting. The tone was off. The language didn’t fit. The message felt generic. By then, budget has been spent and the opportunity has passed.
The core issue isn’t creativity — it’s alignment. Most teams have no reliable way to check whether a piece of content actually matches the persona they’re writing for, before it goes out.
What I Built
Contentify is an LLM-powered application that reads a piece of marketing content and scores how well it aligns with a defined target persona — in real time, before any campaign goes live.
It evaluates content across five dimensions:
- Tone — does it match how this persona prefers to be spoken to?
- Language — is the vocabulary and complexity appropriate?
- Cultural sensitivity — does anything risk alienating the audience?
- Clarity — is the message easy to understand?
- Coherence — does the content hold together as a whole?
Beyond a score, Contentify tells you why something isn’t working and what to change — giving marketers a clear path to improvement rather than just a verdict.
My Contribution
Persona Research & Model Selection
I started by conducting market research and user surveys to build accurate audience personas — understanding not just demographics, but communication preferences, values, and expectations. This data shaped how the model was prompted and evaluated.
After testing several open-source LLMs, I selected Llama 2 for its performance on nuanced language understanding tasks without relying on proprietary APIs.
Application Development
I used LangChain to structure the prompt pipeline, which was the most critical engineering decision in the project. Getting the prompts right — specific enough to produce useful scoring, flexible enough to work across different content types — required significant iteration. Human-in-the-loop evaluation was built in to catch fluency and factual issues the model alone couldn’t reliably flag.
Dashboard & Deployment
The final application was packaged into a Streamlit dashboard and deployed on AWS (EC2 + S3), making it accessible to marketing teams without any technical setup. The interface was designed so a non-technical marketer could get a score and recommendations in under a minute.
Tech Stack
| Area | Tools |
|---|---|
| Language Model | Llama 2 (Meta) |
| LLM Framework | LangChain |
| ML & NLP | HuggingFace Transformers |
| Frontend | Streamlit |
| Cloud Deployment | AWS EC2, S3 |
| Language | Python |
| Project Management | Notion |
Results
Contentify reduced the back-and-forth revision cycle by surfacing alignment issues at the content creation stage — before campaigns launch. Marketers using the tool can iterate on messaging with direct, structured feedback rather than relying on gut feel or post-hoc A/B testing.
What I’d Do Differently
A few things I’d improve with more time or resources:
- More compute — the current deployment runs on a standard EC2 instance. A GPU-backed instance would cut response time significantly, which matters for adoption.
- Larger model — Llama 2 13B or 70B would improve the nuance of scoring, particularly for cultural sensitivity assessments.
- Domain fine-tuning — feeding the model industry-specific content examples would sharpen its relevance for specific sectors (e.g. healthcare vs. e-commerce).
- Browser integration — a Chrome extension that runs Contentify inline while a marketer is writing would be the most natural point of use.
Let’s Talk
If you’re working on an NLP, LLM, or data science problem and want a collaborator who can take it from research to deployment — get in touch.