Use AI to ingest, organize, and retrieve knowledge from meeting transcripts, documents, tickets, email, and team chat — from building a team-wide queryable knowledge base to quickly finding service owners, past decisions, and relevant conversation history. AI accelerates capture and retrieval; the human decides what matters and how it is used.
The engineer decides what knowledge matters, validates retrieved answers and citations, respects permission boundaries, and applies the information in context; AI assists with ingestion, organization, and retrieval but does not replace human judgment or sense-making.
Automatically turn each meeting transcript into a structured Notion page with attendees, decisions, action items, and backlinks to related projects.
Granola / NotionIndex Notion docs and meeting notes into a vector store so a chat interface can answer questions with citations to the original sources.
LlamaIndex / Pinecone / OpenAI APIDeploy a unified internal search experience across Slack, docs, and tickets so users can retrieve relevant context with permissions respected.
GleanI help companies navigate AI adoption -- from strategy to production. Whether you are building your first LLM-powered feature or scaling an agentic system, I can help you get it right.
Design and build LLM-powered products and agentic systems
Go from idea to production with a clear implementation roadmap
Build AI with human-in-the-loop in regulated environments
What's up everyone? Today I wanted to make a quick video since I thought it would be interesting talking about how I currently use AI in my work as a software engineer working for Microsoft in their Azure high performance computing group.
I've been working through where I should hand off to AI versus where I need to stay directly involved—especially in my personal knowledge management. I use Obsidian as my knowledge hub (local markdown notes with links) and Claude Code as my primary agent interface, alongside Notion as an operational hub for tasks/projects. I deliberately keep Claude's access session-based (not always-on) and set boundaries: AI can help organize, search, and triage notes, but it cannot create or modify my "permanent notes," which must be written in my own voice to preserve real understanding. I also experiment with disposable, AI-generated Obsidian vaults for exploratory research (e.g., mapping an industry or concept graph), then selectively turn what I learn into my own fleeting/reference/permanent notes.
Most people use AI coding tools to write code. I'm using Claude Code to help build a personal operating system as a CEO.
I've hooked it up to many different tools that are important to us: Todoist, Slack, Linear, Notion, Salesforce, Gong, email, and so on. I've found it to be far more useful and sticky than the vanilla ChatGPT approach.
I now think of it as a mini coach+chief of staff. Some particularly valuable ways I'm using it right now or things I'm experimenting with:
🧠 Distillation of knowledge
I have a knowledge base that auto-syncs meeting transcripts from Granola, documents from Notion, indexes them by type (1:1s, customer calls, leadership), and makes them queryable across any conversation. It syncs to my Mac hourly. This means I can go see someone at their desk, Granola the conversation, and the context is now available for any future work I want to do with Claude.
🧘♂️ Removing distractions
A small but useful skill I've built is the slack-cleanup skill. It scans my 170+ channels, checks 6 months of activity, cross-references Salesforce and many other tools before deciding what to leave. Keeps me focused.
💭 Self-reflection and persistent memory
It knows our yearly initiatives, quarterly targets, and I write a weekly check-in before the week starts. So when I ask "what should I focus on today?", it doesn't just read my calendar. It checks whether my week is tracking against what actually matters. It's a gentle and useful accountability system.
🧐 Thinking quality
I'm experimenting with a /frame skill that takes messy context and distils it into a one-sentence problem, the binding constraint, and the eigenquestion — the question whose answer determines the answers to all the other questions. I use it to force forward progress on hard problems.
🔎 "Cheap", one-off analyses
For example, a competitive deep-dive into win rates, backed by specific customer quotes on gaps, with Gong snippets. That was a multi-day project compressed into minutes, backed by hard evidence, so I can sense check correctness.
Ultimately, I think the most benefit comes from work that would never have happened in the first place because it was far too expensive to do, vs. making me faster at what I already do. I recommend!