Software that runs
the businesswhile nobody watches it.

One web screen for the head of sales: live revenue by channel, a per-manager leaderboard, activity and presence, a map of every delivery, margins. Pulled from the CRM and the inventory system and reconciled — the numbers finally match.

Replaced a spreadsheet nobody trusted

The same truth, trimmed down to what an owner actually checks, laid out for a phone.

Every sales call over a minute is transcribed and graded on six criteria — interest, discovery, objection handling, call to action, agreements reached. Colour-coded table, per-manager trends, quotes explaining each score, and a Telegram alert when a call falls below a threshold.

~$0.0025 per call · runs on a $2/day cap

Every Monday morning the full sales-meeting deck builds itself — slides and PDF. Revenue by direction, leaderboard, cold-base movement, average checks, upsells, plan versus fact. The owner opens it and runs the meeting.

Was a half-day of prep, now zero

22:00 every day: what actually shipped, split retail / wholesale / regional. Counted from real warehouse documents, not from optimistic CRM stages.

Inside a CRM deal: a branded quote PDF, invoice, waybill and contract, generated from the deal itself and emailed to the client from the responsible manager's own mailbox.

Replaced paid third-party software

Current availability and prices right inside the CRM card, so managers stop messaging the warehouse to ask.

23 competitor sites are read every morning at 09:00. An AI parser normalises wildly different listings to one comparable unit and drops them into a dashboard tab.

Was a person's daily job

Built a wholesale prospect base along a 1,500 km highway corridor from public map data, then cleaned it geographically and deduplicated it.

651 new verified targets · 1,571 wrong ones removed

Finds customers who asked a price and then went quiet, lets an AI filter decide who is worth a nudge, and re-opens the conversation in the same messenger thread in a human voice.

Classified-ad enquiries land in the CRM automatically, with a GPT first reply so nobody waits.

Asks the customer after delivery, collects the rating, and routes anything negative straight to a human before it becomes a public review.

43% response rate

Send the estimate file to a Telegram bot. Two AI passes read it, classify and deduplicate the line items across two languages, then a scheduler lays them out in parallel waves with dependencies, materials and permit tasks inferred.

Hours of planning down to one message

Crews report progress from their phones in plain language; the bot understands it, updates the plan and sends the director a digest morning and evening.

Pins on the floor plan, photos from the phone, finger-drawing markup, and GPT writing the formal defect ticket in English and Russian. Ends in a branded PDF with a near full-page photo per defect.

Replaced a paid SaaS · 193-page report builds in 37s

A procurement module: materials by category, stock levels, prices, low-stock warnings.

CRM numbers into Telegram on a schedule — daily and weekly — read straight from the source of truth rather than recalculated and argued about.

A web tool that scores a property for a flip: ROI, IRR, rental yield, partner split, transaction costs, due-diligence checklist.

A self-hosted search engine builds the vocabulary real people use for a situation — in any language, without paying per query.

Eight social platforms scraped against that vocabulary, normalised into a single shape.

A cheap first pass throws out the noise; a careful second pass has to quote the evidence in the post before it will call something a signal.

Node and SQLite on my own server, a Next.js front end, hard daily spend caps on every paid API, and a nightly quality test that blocks releases when signal quality drops.