Innerboard-Local

`prep` cmd for meeting talking points to go!
`record` cmd to use to let InnerBoard track the terminal usage in the background!
`health` cmd to check the system health of end-to-end workflow functionality ensuring security, performance and reliability!
`list` cmd for previewing your personal vault of reflections and notes!

Inspiration

A 100% offline meeting prep assistant that turns your console history into team/manager-ready updates.

InnerBoard-local analyzes your terminal sessions and produces concise, professional talking points for your next standup or 1:1. Everything runs locally on your machine—no data ever leaves your device.

What it does

Private journaling → structured signals. Extracts key points, blockers, and resources.
Console activity analysis. Turns raw terminal logs into structured “sessions” you can summarize or share.
Actionable micro-advice. Suggests next steps and checklists tied to what you just wrote or ran.
Shareable summaries. Auto-generate weekly reviews, stand-up notes, or personal status updates.
Zero-egress by design. Encrypted local vault; LLM runs via Ollama on localhost.
Modern CLI UX. Rich tables, progress bars, and friendly errors.

How we built it

Local LLMs via Ollama (gpt-oss:20b) with connection pooling and TTL caching.
Security: PBKDF2 (100k iterations), Fernet (AES-128) at rest, SHA-256 integrity checks, strict input validation, loopback-only networking.
Storage: Encrypted SQLite vault with safe file ops and secure deletion.
CLI: Python + Click + Rich; config via python-dotenv; Docker support.
Quality: 50 tests, all green across security, caching, integration, and network safety.

Why gpt-oss

Open weights + local inference: Easy to run and audit offline via Ollama.
Reasoning-first: Handles multi-step extraction from messy shell traces.
Composable prompts: Stable JSON contracts enable “reflection → advice → summary.”

Challenges we ran into

Enforcing true zero-egress while keeping local LLMs fast and responsive.
Normalizing noisy shell logs across OSes and shells.
Designing stable JSON schemas so “reflection → advice → summary” composes reliably.
Making key management secure-by-default yet smooth in daily use.

Accomplishments we’re proud of

A fully offline pipeline with encryption, integrity checks, and strict network isolation.
Structured outputs that drop straight into weekly reviews or stand-ups.
A delightful CLI with rich UI and clear, actionable errors.
Comprehensive tests and a clean developer experience.

What we learned

Local-first changes user behavior: provable privacy → more honest, useful entries.
Terminal traces + prose = a trustworthy timeline of work.
Prompt contracts (schemas + validators) matter as much as model choice.
DX (helpful errors, sane defaults) drives consistent reflection habits.

What’s next for InnerBoard-local

Desktop/Streamlit UI on top of the encrypted vault.
VS Code integration for in-context capture and summaries.
Policy sandbox (MCP-style) and redact-on-export.
Streaming outputs with SSE/WebSockets and backpressure.
Privacy-preserving analytics (all local, opt-in).
One-click installers (brew/winget) and model presets.

Demo (≤3 min)

Record your work: innerboard record … exit (all input/output saved locally with timing).
Automatic chunking: idle > 15 min or > 1000 lines triggers processing.
Local inference: gpt-oss:20b via Ollama (localhost) generates SRE (summary, successes, blockers, resources).
One-command prep: innerboard prep (or --show-sre) outputs Team, Manager, and Recommendations.
Private notes: innerboard add "..." stored in an encrypted local SQLite vault and included in prep.
Privacy proof: zero egress, loopback-only calls; pytest tests/test_no_network.py passes offline.
Category: Best Local Agent (100% offline, encrypted, OSS-powered).