What was built, what it solved, and how it works.

• The persistent home domain.
• Front page for the agentic CRE data system.
• Positioning, methodology, data samples, API and MCP documentation.
• 39 pages, around 23.6K lines, its own Next.js application.

• Application surface for the Python intelligence platform.
• Capital graph, UCC hunt and coverage tools for customers; audit for operations.
• Seven admin dashboards, because human adjudication makes review a product surface.
• 184 routes, with Stripe billing and API key management in the same app.

• Published market surveillance reports.
• Normalised loan and property data in; scored default-risk reports out.
• Scoring returns score, tier and attribution, so a rank can be explained.
• 130K+ loans scored 0 to 100 across 5 signal families, behind approval.

• An LLM client in; loan search, ownership chains and a signal feed out.
• Tool-level access is controlled by a config setting, not code.
• 34 tools built, 15 on the public allowlist, one shared contract.

• 28-agent generative engine optimization platform.
• A brand and category are shared using knowledge base, agents return whether LLMs are citing them.
• Each task has a minimum capability, and the cheapest model clearing it wins.
• $minimum daily cap, and a publish gate blocking anything under 70 on voice.

• Sales call analysis pipeline.
• A recording transcript is added; pain points, costed proposals and a follow-up draft are generated.
• Four sequential Claude phases, each idea citing a quote that proves need.
• 13 routes. Zoom webhooks pull recordings with HMAC signature verification.

Not linkable. It holds real client call recordings. Walk-through on request.

• Chat-driven job search with ghost-job scoring.
• A role requirement given in plain language is scored against postings from three public boards.
• A real Anthropic tool definition inside an agentic loop.
• Query decomposition fans one request into up to 4 short queries.

• Client CRM with proposal and briefing generation.
• Meeting notes are the source for a branded follow-up document with its own URL.
• Price books are embedded, so quotes are never invented.
• 6 live client pages across 4 sectors.

• Resume Researcher, reconfigured for commercial leases.
• A plain-language lease query in; ranked matches are returned.
• Same parse-rank-shortlist-share engine, different domain vocabulary and query fields.
• Domain prompt writes eighteen typed fields, ranking is deterministic TypeScript.

• Domain-agnostic diagnostic engine.
• A metric assessment is scored against cohort-benchmarks, delivering a health score.
• Scoring is entirely deterministic, with the conversational interface being the only model work.
• Built to accept a swappable 220KB knowledge package that supplies metrics, benchmarks and terminology.

• Recruiting database with natural-language candidate search.
• Resumes synced from the system of record; delivering AI-searchable structured profiles.
• Expensive model at ingest, cheap model on every search.
• 14 routes, syncing twice monthly, shortlists shared on one link.

• AI version of a working executive coach.
• Leadership problems and team context get shared with the interface to generate custom-method coaching.
• A knowledge-base overlay across a reasoning model rather than a fine-tune.
• Methodology is content, not code, so revisions are posted without a deploy.

• Lead capture and qualification pipeline.
• A form submission in; an enriched, scored, routed HubSpot record out.
• Error trigger alerts to Discord when an upstream API changes shape.
• 135 workflow nodes across two LLM providers, with a Sheets audit trail.

• LinkedIn prospecting and outreach system.
• An ICP definition generates a sequenced outreach programme.
• A separate worker asks what it may send next, rather than firing on a timer.
• Redeploying the app cannot reset the pacing and burst a day of sends.
• 35 routes. 15 to 20 actions a day, stopping the moment someone replies.

• Multi-channel outreach for contacts already in hand.
• A contact list in; personalised messages out by DM, SMS, email or WhatsApp.
• One model call returns three openers, three pitches and three CTAs.
• Recombining at send time gives 27 message variants from that one call.
• 16 routes, with the channel rules branching inside a single one.

• Signal ingestion hub and scoring engine.
• Events from every other GTM system in; one score per contact out.
• A documented format every system posts into, so nothing arrives twice.
• Contacts match on id, external id, email or LinkedIn URL.
• 41 signal types, 6 categories, 29 tables, each with its own decay curve.

• Deterministic intent scoring engine.
• Behavioural, ad, payment and chat signals in; one ranked call list out.
• No model in it by design, and every score carries its breakdown.
• Four webhook receivers, with anonymous events merging onto a lead on match.

• Meeting booking system.
• Event types and availability can be custom created by the user, who shares it as a public booking page.
• Creates the calendar event and Zoom meeting automatically, timezone-correct.
• 17 routes, in production as the booking page linked here.

• Content publishing pipeline with a human approval gate.
• Generated drafts in over a machine token; approved items out to the blog.
• A separate application, so the review gate is not owned by the generator.
• Items carry a voice score the reviewer can override.

• Interactive prototype of a data-migration flow.
• A target schema and source in; reviewed field mappings and an export out.
• Entirely client-side, no backend, and the confidences are demo values.

• Analyst workspace over the debt corpus.
• A logged-in customer can review maturity wall, portfolio and loan details.
• Shell signs a short-lived token per request, validated before internal routes serve.
• 8 gateway routes, with workspaces provisioning themselves on first call.

Not linkable while the deployed workspace is pointing at a development host. Walk-through on request.

• Operator view onto the GTM signal hub.
• Scored contacts in; pipeline, at-risk accounts, funnel and meetings out.
• Headless systems that runs synthetic data for demonstration purposes.

• Facility operations visibility prototype.
• Headcount, temperature and energy are reviewed to deliver plain-English break-even recommendations.
• Refuses to produce insights until roughly two weeks of data exist.

• Bilingual QR incident reporting for frontline workers.
• A scanned QR code in; a filed injury or breakdown report out.
• Translations are pre-written, ensuring consistency of reporting.

• Reporting surface for generative engine optimization.
• Measurement from the 28-agent platform in; citation scores and content gaps out.
• Currently renders mock data.
• The measurement side is real and runs daily.

• Conditions scoring for surf, ski and beach spots.
• Pulling from six live sources; a decision-engine for best spot to visit today
• 6 sources, 3 activity types, first shipped Nov 2025, Redis-cached.

• Streaming availability and price checker.
• Add your current subscriptions; find out where to watch a show/movie for free, or what it costs to rent.
• Availability and pricing from TMDB across 10 countries.
• Second thing I ever shipped, and still serves today.

• Sixteen repositories that are not on this list.
• Twenty-five repos created in a month, before the Factory existed.
• Among them a waste tracker, a QR generator and a day-camp site built twice.
• 16 cut, 1 rebuilt, and 1 never pushed, so the code is gone.