Hi, I’m Kevin Boller

Inspired by Andrej Karpathy’s “second brain” concept, this is a git-backed system where an LLM handles the organization, thinking, and maintenance of everything I consume and everything I build.

Capture: Readwise Reader is the inbox and ingestion layer for articles, highlights, forwarded emails, and uploaded PDFs, plus Google Drive for working docs. Live external tools (DataCamp, Sigma, Snowflake, Gmail, Calendar) are queried on demand via MCP rather than statically stored, so structured, dynamic data sits alongside all ingested document knowledge.

Compile: Sources land untouched in a raw/ landing zone. Claude and Codex act as an AI compiler, reading each source alongside a hand-authored schema and compiling it into a wiki/ organized by entity type (people, projects, frameworks, sources), one source at a time, cross-linked, with contradictions flagged rather than silently overwritten. No RAG, no vector DB.

Loop: The whole system runs on a compound-engineering loop: Plan (consult the wiki before starting) → Work (do the actual task) → Review (check against what’s already known) → Compound (write new lessons back in). A periodic lint loop separately checks the wiki and index for contradictions, stale claims, and orphaned pages.

Outer loop: A skills/ layer at the repo root holds general operating practices shared across every other project repo I run, personal and consulting alike (including AI-Native House Manager and StreamingLTV), each running its own local Plan/Work/Review/Compound cycle. A lesson learned in any one project hardens the baseline for all the others, so any new project starts at the current baseline of the whole system instead of relearning it from scratch.

Full writeup, including how this connects to enterprise implementations like Stripe’s internal “Kai” system, in the related post.

Built in partnership with Owl & Co (Hernan Lopez, ex-Wondery/Fox/NatGeo) to determine what a streaming subscriber is worth once you combine subscription payments with ad revenue monetization.

The model covers every major US streaming service (Netflix, Disney+, HBO Max, Hulu, ESPN, Peacock, Paramount+, and more), at the grain of individual plan, distributor, and billing cadence.

Tech stack: Snowflake (warehouse, SQL transform layer), Sigma (semantic layer, editable write-back input tables, dashboards), Claude Code and Codex / MCP tooling (AI-assisted data modeling, SQL, and Sigma workbook builds), HTML/JS for client-facing dashboard prototyping.

Key technical work:

• Multi-faceted join topology: Antenna-measured distributors (subscriber, churn, and pricing data) joined with wholesale partners (Comcast, Charter, Verizon), where wholesale volume is modeled through expert input and industry research
• Blended LTV/ARPU formula combining subscriber data and advertising revenue per subscriber, allocated across sports and entertainment inventory
• Sigma write-back tables allowing dynamic model drivers so that updates cascade downstream into workbooks built on elements from the Sigma data model
• External dashboard prototyped in HTML first (faster iteration than building native BI charts), then ported into Sigma
• Deep incorporation of Claude Code and Codex with direct Sigma/Snowflake MCP connections: querying, verifying, and building production data model changes through natural language
• Stable-name materialized view built on top of Sigma’s own materialized snapshot of the model: cut dashboard query latency to 1-2 second cache hits, and gave outside tools governed access independent of Sigma’s unstable internal table names
• Two custom agent skills, streaming-ltv-audit and streaming-ltv-insights, codify the project’s methodology (formula rebuild from raw columns, quality gates, a golden-case regression test), so any AI agent working the model, whether Claude Code or Codex, follows the same rules instead of re-deriving them each session

Why now, the Third Age of Streaming: Hernan Lopez’s research at Owl & Co frames streaming’s evolution in three phases: a first age (pre-2022) that was “all about growing subs,” a second age (2023-2025) that pivoted to profitability with less focus on volume, and a third age (2026+) defined by enterprise value creation through platforming, ingestion, strategic bundling, and game theory. That third age turns LTV into a multi-variate optimization problem rather than a single number:

• Are you close to ARPU parity across ad-free and ad-supported subs (like Peacock), or not yet (like Netflix)?
• How much of your potential revenue are you prepared to share with a distributor, and do you care whether it’s booked gross or net?
• How many of your subs are long-tenured or included in bundles, and therefore less likely to churn?
• What’s the potential for cannibalization?
• Do you want to be a platform or a service, and if a platform, how will you compete for both subscribers and services?
• How are all of these variables impacted by what your competitors do?

StreamingLTV is built to hold all of those variables at once, at the grain of plan, distributor, and billing cadence, rather than collapsing them into a single blended average.

Business impact: Unlocks the “Third Age of Streaming” analysis and LTV publications in Owl & Co’s public Streamonomics newsletter, and is the foundation for an externally facing product targeted to launch in Q4 2026.

Building an AI-native household management app designed for busy parents. Provides households with an intuitive application that learns more about your family and improves over time. Eliminates the setup, development, maintenance, API-cost tracking, and overall burden of building with Claude Cowork/Claude Code, while delivering a comprehensive household management solution that drives actionable workflows rather than piecemeal outputs, like an email digest of the week ahead.

Tech stack: Next.js 16 (App Router), TypeScript, Supabase/Postgres, Claude API, Tailwind CSS + shadcn/ui, deployed on Vercel. Claude API calls run server-side.

Applied AI features:

• Agent email triage: Gmail integration with a per-sender monitoring gate. Claude triages opted-in senders: summarizes, extracts structured data, classifies, drafts replies, and surfaces everything in an Action Needed queue
• Family context layer: Server function assembles a structured prompt block from the DB for Claude API calls. Structured context upfront means outputs are more deterministic and consistent across features
• Skill-based AI behaviors: AI tasks (weekly digest generation, email triage, assistant Q&A) are encoded as versioned skill files that capture domain rules and edge cases. Skills define the expected behavior and are independently improvable without touching application code
• Week-ahead digest: AI-generated “Next Two Weeks” summary card surfaces upcoming health appointments, school events, vendor visits, and tasks. Refreshable on demand
• Draft → Approve → Act pattern: The AI never takes unilateral action. The focus is on keeping humans in the loop before anything is sent or committed
• Telemetry on every AI call: input_tokens, output_tokens, cache_read_tokens, latency_ms, and prompt_version captured per feature at build time. Usage tab shows API cost by day, supporting per-feature cost tracking and prompt regression debugging

Key technical work:

• 8-module app: Dashboard, Email Triage, Kids Health, Car Maintenance, Home Vendors, Utilities, Pets, and a Calendar screen (FullCalendar supporting different school feeds)
• 40+ table Postgres schema in Supabase, including staging/prod split via Supabase CLI with a migration workflow
• Deployed to Vercel with feature-branch workflow

Built a multi-page financial planning dashboard in Streamlit + PostgreSQL to model a 40-year household projection, inspired by FIRE (Financial Independence, Retire Early). Upgraded from a scenario-based Excel financial model to a fully interactive, database-driven application with significantly expanded scenario and sensitivity analysis capabilities.

Key technical work:

• Stood up a 40-year per-account projection engine modeling investment accounts (taxable, tax-deferred, Roth, real estate), variable income schedules, and year-by-year surplus/deficit cash flows routed through a designated buffer account
• Designed a PostgreSQL schema (7 tables) for named scenarios with full assumption, balance, income, expense, and mortgage data. All editable in-app with changes persisted back to the DB, a significant upgrade from the Excel model
• Built a Sensitivity Analysis page with color-coded HTML tables (nominal and inflation-adjusted) showing investable assets at Year 10 across a growth rate × annual expense matrix; each cell calls the projection engine directly to guarantee exact match with the Overview page
• Built a Compare Scenarios page with side-by-side KPI cards with deltas, dual line charts (nominal + inflation-adjusted), and a 10-year side-by-side detail table including inflation-adjusted investable asset columns
• Added an Expenses page with editable budget categories and a Mortgage page backed by a cumulative amortization schedule in Postgres, both with save-to-DB functionality
• Implemented Altair/Vega-Lite charts throughout: assets over time, account breakdown, income vs. expenses, scenario comparison lines, and mortgage amortization
• Full dashboard: 6 pages, projection model, sensitivity engine, and schema. Built iteratively with Claude Code from the Excel model foundation and completed in ~3 hours

Ported a Claude Code-developed, synthetic Netflix streaming analytics HTML dashboard (4 tabs, ~10K subscribers, 105K watch sessions) to a production Sigma workbook, rebuilding the Python data pipeline from scratch in order to generate Sigma-compatible CSVs that were added to and power a Sigma data model.

Key technical work:

• Rewrote sigma_export.py, expanding output from four CSVs to nine; added waterfall.csv, top_titles.csv, top_titles_by_plan.csv, device_summary.csv, and monthly_views.csv as pre-aggregated files for charts that could not be derived cleanly in Sigma from grain-level session data
• Applied all notebook transformations at row level: device watch/completion multipliers, seasonality on watch duration, plan duration/completion multipliers; device session count overrides and monthly seasonality on counts/rates handled in pre-aggregated files
• Added subscription_plan dimension to cohort_retention.csv and waterfall.csv so that the plan filter control on the LTV tab correctly targets all three data sources
• Completed in Sigma: LTV tab (all 4 charts + KPIs with working plan filter) and Content tab (all 5 KPIs + 4 charts across 4 data sources)