I spent years running Partner Success for large brokerages at a real estate CRM company, where my job was adoption across enterprise partners. Since then I've built working tools around the problems I kept seeing from that seat: messy contact books, clients the firm forgets when an agent leaves, referral credit nobody tracks, and health scores that read an account in context. Every one runs on synthetic data and is here to open.
Every number is generated and queried by the system itself, and version-controlled like real software.
The same operating layer the rest of the showcase describes: a persistent system with memory, structure, and a job to do. On this page it's pointed at the problems a brokerage actually has.
A structured library of 1,500+ interlinked notes, plus 8,000+ documents auto-ingested and summarized. The system doesn't just store; it compiles raw input into clean, cross-referenced knowledge, then lints it for gaps.
How it works →60+ custom skills, each a purpose-built AI worker: a morning-briefing analyst, a five-person advisory council, a research synthesizer, a job-application strategist. Invoked with a single command.
Scheduled jobs fire on their own: daily reading gets synthesized and filed, a weekly review compiles itself, daily life capture runs in the background. No prompting required.
Full git version history, a privacy firewall, an audit log on every action, and a single-source-of-truth design so nothing silently drifts.
Each of these started as a question I couldn't answer well enough from the Success seat. Each is a working tool now, built on synthetic books so there is no client data in any of it, and each one is open to read end to end.
A purpose-built Partner Success console for a brokerage's set of accounts. Every account is a structured file system (success plan, stakeholders, decisions, health), and a transparent engine scores each one Healthy, At Risk, or Critical, then reads that score in context, so a ramping account and an established one near renewal get different verdicts, the way mature CS platforms score health. Each comes with the single play that moves it, and clicking Start on any task opens an AI session already loaded with that account's full context. It runs today, ready to point at real partners.
A four-layer customer-data-platform demonstration over a synthetic agent book: entity resolution with a conservative auto-merge band, a referral graph with direct and influence views, cohort analytics, and lead-source attribution, rendered as one interactive page.
Short concept write-ups, each tested on a real corpus before it was written down. Currently: LLM-enriched retrieval that fits a per-user, local-first architecture: explainable search with no embeddings and no pooled index. These pages re-render from their source notes on every site regeneration, so they are always current.
Each is a repeatable agent with a defined job, inputs, and output format. Productized workflows, not one-off prompts.
Pulls calendar, reminders, email, job pipeline, and yesterday's loose ends into one start-of-day report, then drafts the day's note.
Convenes a five-member advisory board (Career, Innovation, Risk, Finance, Marketing) plus a Socratic chairman to pressure-test any real decision.
Compiles the week into a structured retrospective and runs a system health check. Fires itself every Friday, unattended.
Scans a week of inbox and sorts every message into action / job-search / follow-up / FYI / noise so nothing important slips.
Searches the web, filters signal from hype, and delivers a tight briefing on capability shifts that actually matter to my stack.
Pulls an idea from two unrelated fields and cross-pollinates them against what I'm working on. Engineered serendipity.
Runs a half-formed idea through research, a steel-man, and an adversarial case-against, then converges it into a sharp position brief built to be shared, not just filed.
Pulls from the AI-news pipeline and fresh web research to draft a signal-over-hype newsletter issue, staged as a ready-to-send platform draft.
Reads a job posting, assesses fit against my profile, and drafts a tailored application work-up automatically.
Produces sales-grade intel on a target company (stage, culture, partnerships motion, recent news, people), scored against my profile for fit.
Researches the interviewers and the company, picks the right stories to tell, and builds a tight prep sheet plus the questions to ask back.
Researches a hiring manager, recruiter, or referral, then drafts a short, specific, in-voice message tied to one target role and built to get a reply.
Synthesizes a pre-call briefing for any account in the Account OS: health, stakeholders, open escalations, goals, and the one play that moves it.
Takes locally-extracted saved videos (caption, transcript, on-screen text), fact-checks each claim against the web, and scores it for accuracy and relevance to my stack.
Audits deferred and parked threads so nothing promising quietly rots in the backlog.
Writes a structured record of every working session, appends new tasks, and commits the whole system to version history.
A daily unattended pipeline: clipped articles and podcasts are summarized, filed, archived, and committed every morning. No input from me.
Plus the rest of the fleet, from knowledge-graph upkeep to inbox sweeps, model bake-offs, stack triage, and transcript pipelines, working agents that keep the whole system clean and current.
Brokerages adopt what they can trust. So the boring layers are built too: version history, a privacy boundary, an audit log, a single source of truth.
The entire system is a git repository with full history, backed up to a private remote, and the irreplaceable core gets a nightly encrypted off-machine backup on top.
Private journaling is walled off from the cloud AI entirely and processed by a model running locally on my own machine. Sensitive data never leaves the device.
How it works →Every tool the AI touches is logged, with guardrail hooks that can block risky actions. The system watches itself.
Skills are defined once and referenced everywhere, so the system can't silently drift out of sync. A real engineering pattern, applied to personal tooling.
It remembers context, preferences, and history across every conversation; it gets more useful over time instead of starting from zero.
The system keeps a written map of its own architecture and a changelog, re-confirmed on a schedule, so it stays understandable as it grows.
Every number published on this page is recomputed weekly and gated by a drift guard: a stat that silently changes meaning (not just value) blocks the deploy until a human looks. Honest numbers, enforced by machinery.
A watch-list ranker trained on my own ratings, built like an ML project should be: baseline bake-offs, an ablation showing a quarter of the corpus ranked just as well as all of it, and a leakage audit that banned my own 'vibes' tags as features: leaky at training time, unavailable at recommendation time. The write-up of what leaked and why is on the writing lane.
A multi-pass pipeline (research, steelman, adversarial case-against, converge) that turns a half-baked idea into a publishable position document. Separates an idea's merit from my ability to personally build it, so a strong idea I can't ship still becomes a sharp document worth sharing.
An allowlist-driven snapshot of the knowledge base, derivative-only, publishes to a private GitHub repo on demand. The artifact people get when they want to see how I actually think: not slides, the raw working corpus minus the private bits.
A dedicated always-on Mac mini to host the long-running services (podcast ingestion, persistent assistant host, knowledge-graph MCP) so the laptop can sleep without breaking the system. Architecture decided, hardware pending.
A periodic publication built on the idea-brief pipeline's output. The first pieces are already live on the writing lane; what remains is the recurring send. Each brief doubles as portfolio and audience: no double work.
Apple Watch trends sanitized into a thin derivative the local AI can read (sleep, exertion, recovery) so its suggestions can factor in whether I'm sharp or fried. Strict boundary: locally-derived summaries only, never raw health data, never leaves the machine.
Adoption was my lever in Partner Success. These tools are what happens when that instinct gets a build system behind it: the problem named from experience, the solution built and tested, the evidence published.