AI Workflow Engineer
Prompt engineering · Workflow automation · System integration
I design and ship Claude-powered automations for real business operations — turning repetitive customer-service and sales work into fast, reliable, measurable workflows.
About
I'm Kitz Rulona, a Computer Engineer and founder of Hex Colony, a tech startup that automates customer-facing and back-office operations. I use Claude Code every day and build with the Claude Messages API — system-prompt design, structured outputs, few-shot prompting, and evals.
My background spans software, APIs, and IoT/firmware, so I'm comfortable from the prompt all the way down to the integration. I care about workflows that are real and measurable, not demos that only work in a screenshot — and I keep a human in the loop wherever the stakes are high.
Skills
Grouped the way this role asks for it — from the prompt to the integration.
Certifications
Anthropic-issued Claude credentials, plus a HubSpot CRM certification.
Selected work
Real, runnable prototypes spanning the whole role — customer service, sales, internal operations, prompt optimisation, and an autonomous agent on a custom MCP server. Each has engineered prompts, schema-enforced output, a human-in-the-loop gate, and evals or tests that measure (not guess) quality.
Classifies inbound service messages (booking / quote / parts / complaint), scores urgency, drafts an on-brand reply, and logs to a CRM — flagging complaints and safety issues for a human.
▲ Escalation accuracy 91% → 100% via eval-driven tuning Read the case study →Qualifies inbound leads on a BANT rubric, assigns hot/warm/cold tiers, drafts personalised follow-ups, and prints a prioritised call queue to the CRM.
▲ Escalation accuracy 75% → 100%, tiering held at 100% Read the case study →A knowledge-grounded assistant for service-desk staff: answers from the dealership's own materials, cites the source document, escalates when it's not covered — and has an onboarding mode that trains new hires.
▲ Grounded & cited · refuses to invent policy Read the case study →Runs competing prompt versions through labelled test cases, scores them with deterministic checks plus an LLM judge, and produces a before/after comparison — proving a prompt change actually improved performance.
▲ Measured a prompt change: 88% → 92% accuracy Read the case study →A custom MCP server exposes the service desk's CRM and knowledge base as tools; a Claude Agent SDK loop triages the overnight message queue on a schedule — matching customers, grounding prices, drafting replies, and flagging complaints and safety issues for a human. TypeScript.
▲ Custom MCP server · Agent SDK · scheduled · MCP tools verified end-to-end Read the case study →A structured interview becomes persistent memory; public content from three chosen experts becomes a raw-notes → wiki knowledge pipeline; a custom Claude Code skill convenes them in their own voices on any decision — plus a watchdog that emails the board's letter when I go dark. I was my own first client; I use it daily.
▲ Live daily use · open-source template · drift watchdog verified end-to-end Read the case study →Résumé
AI/automation and prompt-engineering first, with the metrics behind each project.
Contact
Available ~20 hrs/week, remote, PHT — flexible across timezones.