SHAWRANA

Design&Rank operations system

SR-001 · 2015 · CL-A · BUSINESS SYSTEMS (PRE-AI)

Design&Rank started on paper. Leads on paper, closers dialing from paper, customer service logging jobs on paper for the dev team.

I picked a modular base and rebuilt it with my dev team around the whole company: SDRs, closers, customer service, retention, fulfilment and monthly billing in one pipeline.

By the end, every part of the company ran through it, billing and fulfilment included.

40 · people on one pipeline

Prime CS dialer platform

SR-002 · ~10 yrs running · CL-A · BUSINESS SYSTEMS (PRE-AI)

Prime CS sets appointments for car dealerships and gets paid per lead, so every extra minute per lead eats margin. Before the build it ran on spreadsheets and manual dialing.

I designed a dialer pipeline with voicemail detection, so agents only ever speak to a live person, with the right record already on screen.

It handles every client's intake and delivery format, from email to SFTP, and gives supervisors dashboards to run each campaign to goal.

~10 yrs · in production

Augmented underwriting

SR-003 · AI era · CL-B · AI IN PRODUCTION

Built solo for a micro-lending company. Human underwriters stay in charge; AI adds decision signals to every application.

The data is processed on the company's own box with local models, so applicant files never leave it.

Approvals and declines got faster, and fewer loans defaulted.

Saivpoint save desk

SR-004 · 2026 · CL-B · AI IN PRODUCTION

An AI desk that answers refund, cancel and billing messages for a digital-offer subscription brand I run.

It handles the routine cases and hands anything unusual to a person.

83% · resolved by the AI alone

Company rolodex

SR-005 · 2026 · CL-B · AI IN PRODUCTION

One registry and app for every company I own or manage: directors, fiscal year-ends, deadlines, incoming letters and a ledger.

An agent works from it with fixed rules. Secrets stay in a password store; the database only keeps pointers.

33 · companies run from one app

Affiliate review engine

SR-025 · 2026 · CL-B · AI IN PRODUCTION

A review site for paid online communities, run by an agent pipeline. Agents gather the evidence, draft each review and recheck every price daily.

Commission can set list order. It never changes a score or a verdict.

A build check blocks any page whose title or verdict disagrees with the data. It caught a live wrong price on its first run.

107 · reviews live in week one

AI call desk

SR-006 · 2026 · CL-C · VOICE + OUTBOUND AI

It places real phone calls for me: parts desks, suppliers, quotes. One brief in, one call out, then a transcript and a structured result on my phone.

It runs a speech-to-speech model over a Canadian carrier, handles French, answers callbacks, and hangs up on voicemail.

Every call needs my go first.

Collections outreach

SR-007 · AI era · CL-C · VOICE + OUTBOUND AI

A context-driven SMS and email agent for the micro-lending company that reaches defaulted clients and sets up payment plans.

A compliance check runs against every outbound message before it is sent.

Conversational TV guide

SR-027 · 2026 · CL-C · VOICE + OUTBOUND AI

Ask the TV for "something for the kids, they're 6 and 9" and it reads the live guide of your own IPTV service, then offers a short list to play. Voice input from a phone remote is in build.

No generative model in the loop: every decision is a typed call to Jev, a decision model that answers in about 300 ms.

Against a cheap generative model on the same 74 cases, it scored 74 to 70 and ran 64x faster. The generative model served a show above a child's age limit; Jev never did.

It is a client only. It never hosts, sells or touches a stream.

80/80 · eval cases, zero age-limit misses

Marketing feedback system

SR-009 · AI era · CL-D · INTELLIGENCE PIPELINES

Reporting across Google Ads and Meta for the micro-lending company, fed back into creative and ad strategy.

An LLM compliance layer reviews all marketing output before it runs.

Evergreen wiki

SR-010 · 2026 · CL-D · INTELLIGENCE PIPELINES

Every agent session I run gets mined each night. Learnings land in a wiki with backlinks, a health check, and an index the next session reads first.

It is how a lesson from one project reaches the others without me repeating it.

100+ · pages, grown from sessions

Geo-leak lead finder

SR-011 · 2026 · CL-D · INTELLIGENCE PIPELINES

Finds businesses whose Meta ads spill across a border they do not serve. That leak is a warm lead.

A pre-triage classifier cuts paid lookups before anything is scored.

12,800 · ads screened by one pipeline

Search + AI-answer engine

SR-026 · 2026 · CL-D · INTELLIGENCE PIPELINES

The SEO system behind the review site: Google sweeps, AI Overview and Perplexity citation probes, and search engines pinged on every deploy.

Every page change gets a ledger row, a 28-day baseline and a control set. Nothing counts as a win until a six-week window closes.

It runs as one skill whose rules change when the evidence does. This site runs on it too.

428 · queries swept for $1.64

The agent fleet

SR-012 · 2026 · CL-E · AGENT INFRASTRUCTURE

A strong model drives: it plans, judges and writes the final word. Cheap, uncapped lanes do the typing and the browsing. A read-only scout does volume retrieval. A reviewer with fresh context checks the work.

Claude keeps the planning and the review. The typing goes to the cheapest model that passes a bake-off: same harness, same 34-check verifier, only the model changed. The winner scored 34 of 34 in 69 seconds; the runner-up took 189.

2.7x · faster at the same score

Telegram bridge

SR-013 · 2026 · CL-E · AGENT INFRASTRUCTURE

Full Claude Code from my phone, built on an open-source bridge and extended: one private group per project, voice notes transcribed, a scheduler for recurring jobs.

Anything longer than a couple of minutes goes to a background lane, so the chat stays free while the job runs.

Approvals only apply to the message I reply to.

Seeded-defect QA

SR-014 · 2026 · CL-E · AGENT INFRASTRUCTURE

A builder agent does not get to grade its own gates. A 13-agent workflow built each quality gate, then attacked it with a planted defect it should catch.

A gate that misses its own seeded defect is fixed or marked not trusted.

VinSight market engine

SR-008 · 2025 · CL-F · VINSIGHT DEALER INTELLIGENCE

The data spine for a dealer-to-dealer car business. Wholesale auction results, retail listings and dealer inventory land in one database.

Every car gets a VIN decode and a canonical trim, so a comparable is the same car and not only the same model.

Comps are tiered, weighted by recency and adjusted for kilometres and condition.

280K+ · wholesale sales behind the comps

The deal gate

SR-019 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

The pricing engine shortlists deals. A strong model then reads each one in full and returns BUY, NEGOTIATE or PASS with a walk-away price.

Its validation run audited the engine: the top deals claimed $262K of margin, and the gate's realistic estimate was a $20K loss.

Borderline scores get three independent runs. When the runs disagree, that disagreement is the signal.

$262K · of phantom margin caught

Dealer-site crawler

SR-020 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

Dealer websites publish the VIN and full specs that marketplaces hide. I surveyed 742 dealer sites and fingerprinted the website platforms most of them run on.

One extractor per platform reads each dealer's used inventory, deduplicated by VIN.

509 · dealer sites crawled

Cost-basis X-ray

SR-021 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

It matches cars on dealer lots against the auction where the dealer bought them: hammer price, date, kilometres and condition.

Every negotiation starts from the seller's real cost. It also catches private sellers who are dealers in disguise.

$4,495 · average dealer markup, measured

High-end radar

SR-022 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

Every car listed between $150K and $600K in Canada on one board: dealers, marketplaces and private sellers.

Rules set the trim first. A model fills the gaps and returns nothing when unsure, because a confident wrong trim poisons every comp.

Trim coverage went from 41% to 72%.

4,332 · cars over $150K tracked

Dealer brokerage desk

SR-023 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

I post a car in a chat. The agent runs comps by province, emails same-make franchise dealers, follows up and negotiates inside my price band.

An AI receptionist answers the desk's phone line. Accepting an offer still needs my go.

The minimum price is enforced in code: it never enters a prompt, and any quote below it is dropped.

Dealer back office

SR-024 · 2026 · CL-F · VINSIGHT DEALER INTELLIGENCE

Bills of sale, registrations, invoices, funding and a deal tracker for the dealer companies, in one tool.

Each car moves through chained buy and sell legs, with wire, registration and document status on every leg. Most forms fill themselves from the VIN.

The ledger matches bank deposits to invoices, and a person confirms every match.

Lego system

SR-015 · 2026 · CL-G · OFF-DUTY BUILDS

Sort by shape, never by colour: eight labelled bins. Next comes an inventory of every part, then custom instruction booklets designed only from parts already owned.

Designing around the inventory is what saves the money.

Jeep WK2 parts hunt

SR-016 · 2026 · CL-G · OFF-DUTY BUILDS

A scheduled agent hunts used body panels for a 2014 Grand Cherokee: checks fitment, emails yards, logs quotes and replies.

It negotiates. Nothing gets bought without my confirm.

Library-hold sweeper

SR-018 · 2026 · CL-G · OFF-DUTY BUILDS

Three small command-line tools: check my holds, sweep what is on the shelf at my branch right now, and place a hold.

The catalogue API signs every request, so the sweeper implements the signature instead of driving a browser.