Real systems. Real clients. Real numbers.

Problem

6.5 full-time operators were tied up on manual goods receiving and reconciling two accounting contours (management vs accounting). Receivables tracking was reactive, and shrinkage was being absorbed rather than flagged.

Solution

Mobile receiving app that scans barcodes directly into the ERP, an auto-reconciliation engine between the two accounting systems, and an anomaly scoring layer for fraud, shrinkage and receivables risk. A Telegram bot lets managers query the system in natural language.

Problem

Support tickets were piling up faster than the team could clear them. Roughly a third of inbound questions were repeats of things already answered, and response quality varied wildly between agents.

Solution

Analysed 1000+ historical tickets for classification, sentiment and resolution time, built a RAG layer over the knowledge base, and shipped a Telegram L1 bot for FAQ, payment statuses and common troubleshooting. A scoring layer compares agent answers to the best-in-class response pattern.

Problem

A new association needed a platform to actually run tournaments — scoring, handicaps, leaderboards — plus a web presence that could carry a federation-grade brand. Nothing off the shelf covered Stableford, Scramble and Match Play in one place.

Solution

Shipped a Next.js platform on gpaga.ge with Google OAuth, a mobile hole-by-hole carousel, a desktop scorecard table and leaderboards by division. All four scoring formats (Stableford, Stroke Play, Match Play, Scramble) are supported natively, with handicap-aware Stableford point display.

Problem

Ideas, decisions and commitments were being generated in group voice chats and evaporating within a day. No archive, no action-item extraction, no way to search months back.

Solution

A Telegram bot catches audio, text and YouTube links, transcribes with Deepgram Nova-3 including speaker diarization, and runs a Claude Haiku pass that returns a summary, extracted ideas and commitments straight back to the chat. Files are written into Obsidian via GitHub sync for long-term retrieval.

Problem

Telegram and several other call platforms do not expose a record button. To feed the transcription pipeline we needed a reliable way to capture both microphone and system output on macOS, without relying on browser extensions or screen-recording hacks.

Solution

A native macOS app (Swift + AVFoundation + Core Audio Taps) with a floating REC overlay and a menu-bar status item. Records microphone and system output as separate tracks and uploads them to the processing backend once the session ends.

Problem

Running an AI-native consulting operation means fifteen-plus subscriptions and as many API keys, spread across Gmail receipts and vendor dashboards. No single place to see what is actually being spent each month or catch forgotten renewals.

Solution

A Chrome extension plus a local FastAPI + SQLite backend. A three-step Gmail pipeline — Discover → Classify → Scan — finds billing domains, then pulls only confirmed receipts. Fuzzy matching on sender names and billing patterns keeps false positives low.