GoReach

The problem

LinkedIn content tools generate generic AI slop. The output sounds nothing like the person posting it, and most teams either give up after a week or end up rewriting everything anyway. The real problem isn't generation. It's matching voice. And matching voice is a retrieval problem before it's a generation problem.

The approach

GoReach treats every user's past writing as the source of truth. The system ingests your writing samples, builds a personal RAG index over them, and uses that index to ground every generation in your actual phrasing, sentence rhythm, and topic preferences. Prompts are layered: a base writer prompt, a voice-matching prompt grounded in retrieved samples, and a topic-specific prompt for the post angle.

The full stack is mine end-to-end: Next.js on the front, FastAPI on the back, AWS for infra, GitHub Actions for CI/CD with Docker. CI/CD ships changes to production on every merge to main with no manual steps.

Tech decisions worth noting

• FastAPI over Next.js API routes for the AI layer. Easier to scale the inference and retrieval workloads independently of the UI.
• RAG over fine-tuning for voice matching. Cheaper to update, easier to debug, and the retrieved context is visible in the output for trust.
• Self-hosted on AWS rather than a managed AI platform. Full control over latency, cost, and data flow.

Outcome

[TODO: Neeraj] Traction metrics. Sign-ups, posts generated, active users, retention, etc. Even early numbers are fine: "live since July 2025, X users, Y posts generated".

What I learned

The hardest part of building a personalisation product is not the model. It's the cold-start problem. Users land with no writing samples and want quality output immediately. We solved it with a guided "writing sample" onboarding that asks for three short paragraphs in their voice. The flow took longer to design than the rest of the generation pipeline.