I build pipelines that turn operational data nobody reads — call recordings, CRM exports, ad-platform logs, Search Console dumps — into decisions someone can act on. Transcription and classification with LLMs where judgement is needed, deterministic checks in code everywhere else, and a person at the gates that matter.
The proving ground is a multi-location US home-service operator (appliance repair, ~1,600 jobs a month, eight metros). Everything below was built there first, on live data, and then generalised.
Second track: GEO / AI visibility and local SEO — getting a business cited by ChatGPT, Gemini, Perplexity and Google AI Overviews, and measuring whether that actually moves anything.
| 📞 call-audit-demo | Folder of call recordings → local Whisper transcription → evidence-backed LLM classification against a fixed outcome taxonomy → sortable report. Answers "where did this conversation lose the booking?" Under $5 per 100 calls. Anonymised version of a production pipeline. |
| ✅ article-qa-pipeline | Verification harness for AI-assisted content: deterministic checks in code, editorial judgement split into isolated single-criterion agents, a human at three explicit gates. Built by reading a real production session and separating judgement from unwritten spec. |
| 📊 refactor-dash-forked | Google Local Services Ads dashboard across multiple locations. A scheduled Playwright collector pulls daily spend, leads and cost per lead; a React front end renders trends, geo comparison and efficiency views. · live |
| 🔎 CanniScope | Free browser-based keyword-cannibalization auditor. Reads a Search Console export, finds pages competing for the same service + geo, ranks clusters by risk and recommends merge / 301 / canonical / leave. · live |
| 🧠 site-architecture-mri | Crawls a site, clusters pages with local sentence embeddings, computes internal PageRank from the real link graph, and reports orphan topics, weak clusters and under-linked money pages with concrete "link from here" suggestions. |
mobile-portfolio — three shipped React Native + Expo + TypeScript apps with live web demos, unit-tested logic and GitHub Actions CI: Client Hub (phone-first field-service CRM), Eat2Beat (calorie and macro tracker), Solva (storefront). · live demos ▶
One anonymised multi-location client. Numbers as of August 2026; sources and windows are named on odariuk.com.
- 97,800 call records matched to CRM jobs, transcribed and classified
- 11,724 jobs re-attributed to their real source after joining call logs to the CRM
- ~$23K in invalid Local Services Ads leads identified and disputed
- 63 → 2,605 AI answers citing the site in six weeks across 7 engines (ChatGPT alone 1 → 813, Jul 4 → Aug 18, 2026, Ahrefs Brand Radar)
- 6,119 → 214,496 monthly search impressions, Sept 2025 → Aug 2026 (Search Console API)
- 64 → 496 indexed pages after a 400+ URL architecture rebuild
Open to fixed-scope projects and remote engineering roles — async, judged on outcomes. Details and pricing at odariuk.com · odaryukigor@gmail.com