TL Drafter
An author-focused drafting tool that replicates a CEO's voice by anchoring generation on past writing samples and A/B variant comparisons.
The Voice Problem
Executives spend 4 to 8 hours per article. AI tools and ghostwriters both fail at the same thing: voice. The output sounds smart but generic. The tool I worked on was built to fix that for the Eskwelabs CEO, using his past writing as the anchor.
Phase 2: The Three Pieces
I inherited a Phase 1 MVP and owned Phase 2. Three pieces. The writing samples library, where users upload PDFs of past articles and the system extracts text and ties them to a writing session. The A/B comparison mode, which generates two variants side-by-side using a single LLM call with delimited output to keep cost and latency down. And the AI pipeline refactor, which made the system stable enough to handle three to five full articles injected into the prompt as style references.
Architectural Bet
We skipped RAG and embeddings entirely. The bet was that voice replication isn't really a retrieval problem. It's about giving the model enough connected, in-context style data to pattern-match the rhythm and word choice. Modern context windows are big enough that this works. The A/B mode made the difference visible. Variants with samples sounded like the author. Variants without sounded like AI.