Cover Letter Local
Open-source app that writes evidence-backed cover letters from your CV, LinkedIn and GitHub.
- Built during my AI internship at Microsoft
- Open source · MIT
- Web app · runs locally
- Jul – Aug 2026
- Live
- 38K
- 204
- 208
What it is
Cover Letter Local is an open-source app that writes cover letters from your own evidence: your CV, LinkedIn and GitHub. It checks every claim in the draft against that evidence.
I defined and built it alone during my AI internship at Microsoft in July and August 2026. It is an independent internship project, not a Microsoft product.
The problem it solves
Tailoring a cover letter for every application means explaining the same background again, researching the company and still avoiding claims you cannot support. Generic AI tools write fluent text but invent details.
What I built
The app reads your CV, LinkedIn and GitHub, researches the company with six parallel agents and verifies each claim in the generated letter against its evidence.
Retrieval is hybrid: dense search plus BM25, fused with RRF and optionally reranked. A multi-provider LLM gateway supports Foundry Local, Ollama, Azure OpenAI, OpenAI, Anthropic and Gemini.
Evidence-based profile
CV, LinkedIn and GitHub imports end in a review step. Nothing is merged until you approve it.
Your writing voice
Past letters are analysed into a voice fingerprint and a writing playbook.
Company research
Six parallel agents research the company: profile, culture, tech stack, recent signals, interview context and the role.
Claim verification
Each concrete claim in the draft is marked supported, partly supported or unsupported.
Local first
Runs on Microsoft Foundry Local or Ollama by default. Cloud models are used only if you choose them.
Export
PDF and Word export are rendered locally.
How it works
Import your CV, LinkedIn and GitHub and approve the proposed profile.
Learn your voice from earlier letters.
Add a job posting. Six agents research the company and a fit analysis scores the match.
Hybrid retrieval (dense search and BM25, fused with Reciprocal Rank Fusion, optional reranking) picks the evidence.
The model streams the letter, then each claim is verified against your profile before you export.
The result
About 38K lines of code, 204 commits and 208 backend tests, released as open source under the MIT licence.
What I learned
- Before this project I had not built a RAG pipeline or used embeddings or Foundry Local. Building it end to end taught me why lexical and semantic retrieval solve different problems and why they are better fused than chosen between.
- Verifying claims against evidence changed how I design AI features: the model writes, but the product decides what is allowed to be said.