What we build, with the proof.
Each line names what WYTOR builds and the evidence that it has built it. No capability is listed without an example that exists.
AI agents and agentic systems
Agents that take real actions inside a business: calling, drafting, classifying, filing, with approval gates where money or customers are involved.
Evidence
Celabe's AI concierge phones vendors, collects itemised quotes and hands every booking to a human planner before anything is confirmed. WYTOR's own engineering runs on agents that open pull requests for a person to merge.
Voice AI pipelines
Speech in, action out, at production latency: telephony, streaming recognition, model orchestration, and the measurement to keep it fast.
Evidence
Published at ICANN 2026: a systems approach that cut median end-to-end latency by 41.8 percent across pipelines serving more than 40,000 locations.
Knowledge systems and retrieval
Making a company's documents, tickets and code answerable: ingestion, chunking, retrieval, citations, and the evaluation that proves the answers are right.
Evidence
Skillfile distils a repository into the context an AI engineer needs and serves it over MCP; it is being proven on WYTOR's own codebases before release.
LLM evaluation and observability
Test sets, severity scoring, regression gates and dashboards, so a model change is a measured decision rather than a feeling.
Evidence
Research under review on scoring language-model failures by severity; evaluation harnesses built for every WYTOR product that calls a model.
Mobile products
iOS and Android apps built once with Expo and React Native: sign-in, payments, notifications, offline caching, store release and the updates after.
Evidence
Seven consumer apps taken to the App Store and Google Play in one stretch, with in-app purchases, health integrations and multi-language support.
Backend infrastructure and integrations
APIs, databases, queues, scheduled jobs and the glue between a legacy system and a modern one, run with backups, monitoring and a runbook.
Evidence
A fleet of self-hosted Postgres and Supabase stacks on a single server, nightly backups with restore drills, and an outside monitor that checks every stack every five minutes.
Web platforms and automation
Sites, dashboards and content pipelines on Cloudflare and Vercel: Workers, durable workflows, object storage, edge caching, cookieless analytics.
Evidence
This site. A content pipeline that renders and publishes on a schedule from a server WYTOR runs. Image delivery moved behind a CDN after a load test showed where the ceiling was.