LocalGist

Light Mode
Dark Mode

Inspiration - Interviews, meetings, and user research stack up as transcripts, and the usual way to analyze them is to paste confidential conversations into a cloud API. That's a non‑starter for sensitive material. I wanted LLM‑grade transcript analysis with none of the privacy trade‑off, running entirely on your own machine.

What it does - LocalGist is a windows desktop app that reads your .txt/.md transcripts and answers questions about them with structured, quote‑backed results:

A synthesized overview of what the sources say
Key findings, each citing the exact supporting quote and its source
Recurring themes across the selected transcripts
Suggested follow‑up questions Four modes (Decision brief, Themes & patterns, Risks & objections, Customer language) reframe the same sources for different jobs. No data ever leaves your computer.

How we built it -

A hardened Electron shell (context isolation, sandbox, no Node in the renderer) hosting a local Node HTTP server + JSON API.
OpenAI's gpt-oss open model, run locally via Ollama(Ollama cloud is also available), for synthesis. LocalGist auto‑detects installed models and lets you pick one.
A grounded & extractive fallback (keyword + question‑overlap sentence ranking) so it still works with no model present.
A vanilla HTML/CSS/JS UI with system‑aware light/dark themes.
Packaged as an NSIS installer with electron‑builder.

Challenges we ran into - Keeping generation honest: prompting gpt-oss to cite only exact quotes from the supplied transcripts and return strict JSON, with a parser that degrades gracefully. Designing a secure Electron architecture where the renderer has zero direct file or network access. Staying useful fully offline.

Accomplishments that we're proud of - 100% local private by-design analysis with real citations, useful with or without a model & clean, focused interface.

What we learned - Running OpenAI's open models locally is now genuinely practical for private, on‑device AI, and grounding every finding in a real quote is what makes LLM output trustworthy.

What's next - macOS and Linux build, whole folder and larger corpus analysis, export to Markdown/PDF, and speaker‑aware parsing.