Raj Ranpariya

A presentation tool for comparing heating and cooling systems on the three things a customer actually feels: comfort, efficiency and sound. One self-contained HTML file for a tablet, with embedded photos, drawings and unit sound demos calibrated so the differences you hear are proportionally real, a comfort-and-performance ladder that builds as you add systems, plain-language explainers on right-sizing, humidity and warranties, and a one-tap PDF summary the customer can take away. Built by a salesperson for the sales conversation; it works offline and travels by AirDrop.

A single-purpose iPad app for the sales appointment, built for my own day job. It replaces the paper comfort-survey booklet with a seven-step flow: the job, a 35-question comfort survey, an observation sheet for the existing equipment and ductwork, typed tape-measure dimensions, a live rotatable 3D model built from those numbers, a 16-shot photo set and an optional signature. It exports one self-contained crew packet the install team can review remotely, which removes the separate pre-install walkthrough. It runs fully offline with no backend, and it has two honesty rules I refuse to improve away: photos never generate the 3D, and a blank field stays visibly blank rather than being guessed.

Backorder Buster, a small automation for my day job that attacks the most expensive surprise in a sales pipeline: a unit that is sold, scheduled and then turns out to be backordered. It signs in to the distributor portal with a captured session, checks the availability of every model on the upcoming installs, classifies each one as in stock, low stock, backordered or restricted, and emails a short report before the install date is at risk, so a model can be swapped or a customer warned while there is still time. Built because I was the person who had to make that phone call.

Voice-to-structured-action layer for a field-service operations system. A seven-assistant squad: one call-taker greets, withholds the caller's name until confirmed, triages urgency from a per-trade knowledge pack, and hands off to six specialists. The voice platform's 30-second budget per tool call forced a split: anything on the conversational path runs in-process against the database; slower work falls to a workflow engine behind a 25-second abort. Caller identity comes from telephony metadata and overwrites whatever the model supplies; a booking exists only when a row is written. A weekly text-simulation harness runs three scripted scenarios across 18 trades; the latest run passed 18 of 18.

Scheduling and work-order agents. One workflow-engine agent books, reschedules and cancels appointments against existing bookings; the other turns a won job into a work order that a manager must review before the scheduling and billing handoffs fire. Every request passes authentication, a role check and per-tenant task toggles; installation, inspection, callback and emergency bookings are held for human approval. The voice channel offers slots filtered against existing appointments and has a review-first mode for busy periods. A daily automated check exercises 14 agents' write paths and asserts a row persisted; the last seven runs passed 14 of 14. Dispatch recommendations remain stubs.

Permit-and-inspection lifecycle agent. One webhook entry point routes six actions behind authentication and role checks; permit status is an explicit state machine that spans inspection scheduling and results, unknown actions return 400, and a locked design rule says nothing is ever auto-submitted. A headless-browser sidecar, memory-capped and loopback-only, with form filling, CAPTCHA detection and a WebSocket hand-off to a human, is built and health-checked, but no workflow calls it yet and only test screenshots exist. What runs today is record creation plus a deterministic permit-required decision, exercised by a daily automated check.

Post-job follow-up agents, the pair I am most pleased with because the timing is the product. An accounting agent creates invoices, reports collections, and exposes four payment-exception types through a manager-only action. A review-and-referral agent is triggered by an invoice-paid-in-full event or a daily scan of completed work; a per-trade playbook supplies the ask timing, and suppression rules cap requests at two attempts with a 48-hour minimum gap, logging the reason for every suppression. Every complaint type is flagged for a human; none auto-resolves. A composite sequence is designed to wait 48 hours before the review ask and seven days before the referral touch; that sequence has not yet run.

Two safety layers for unattended operation. A content pipeline where a scheduled writer drafts and self-scores short posts, then submits them to a deterministic code gate (blocklist, banned claims, a stat whitelist with inline sourcing, no personal attribution, a quality floor) before anything publishes; a one-call kill switch returns it to human approval. Beneath it, a separate outbound guard for email: two hardcoded invariants no configuration can weaken (one forbids contact with a restricted party), HMAC clearances the transport demands, an append-only ledger, and a post-send sentinel. 119 tests; two independent adversarial review rounds found 25 issues before deployment.

Verification I trust more than reports. A green test suite that shares a predicate with the code cannot falsify it, so I delete each gate in turn and watch which tests still pass; on the email guard that found four sub-checks that could be removed with zero failures, and they now assert behaviour. A daily smoke check fires each agent's main write and asserts a real row landed, not a success flag. A weekly voice harness replays scripted callers across 18 trades with an independent judge, recalibrated once after it scored a safety-first emergency redirect as a failure. Done means an external effect a separate checker verified.

Deterministic scripts score a five-component market regime daily from free public data, apply a long-term trend filter to a watchlist, snapshot a paper portfolio, and write a graded weekly review to my phone. Anything touching real money executes only after an approval reply from me; an unanswered request expires after two hours and places nothing. A file-based kill switch is checked at every point on the order path, and the self-heal loop may restart monitoring jobs but never the money path.

Source documents and the actual question in; a checklist, a comparison table, or a list of questions to ask a professional out; every line reviewed by me before anything is filed or chosen. The value is not the generated text. It is shortening the distance between an unfamiliar problem and a reviewed, concrete next step.

Nothing clinical, no numbers to chase; just a consistent plan that survives contact with a real schedule.