AI & Automation
Operational AI, not a POC.
I deploy AI inside real companies: business process automation, targeted prompt engineering, Vision & Voice model integration. Everything below runs in production.
My approach always starts with the cost of a process before adding AI. The goal is not the tech, it's measurable time savings and a tool the team can keep maintaining.
For AI applied to engineering itself — agents, workflows, CI/CD — see the Agents & tooling card.
Real cases
~92h/month saved · call-back time halved
- End-to-end AI-driven sales pipeline: automatic lead enrichment, smart routing (2 setters / 3 closers).
- Pre-call personalized messages and real-time steering dashboards.
Stack Make · n8n · GoHighLevel · Pabbly · Supabase · Next.js
90% of property reports auto-filled · built with BMAD
- Web & Mobile app (offline-first React Native) automating real-estate inspections: AI detection of materials, colors and wear.
- Built with agent-driven agile (BMAD method), from PRD to implementation.
Stack BMAD · AI Vision · Voice · Next.js · React Native · multi-tenant SaaS
ATS score computed, not hallucinated · verifiable AI judgments (23 → 41 of 48 criteria)
- The LLM extracts and normalizes keywords from both the profile and the job posting; the ATS score is then computed mathematically on the real intersection. No hallucination, a reproducible rate.
- Public scanner backed by verifiable AI judgments (JEV / TypeSafe AI): every verdict carries a probability and an excerpt actually found in the CV, and the system abstains when evidence is thin. Coverage went from 23 to 41 of 48 criteria after one iteration on text preparation.
- Developed with agent-driven agile (BMAD method), with PM, architect and dev agents orchestrated end-to-end.
Stack BMAD · JEV (TypeSafe AI) · Next.js 16 · React 19 · Supabase · AI SDK
Product AI stack
n8nMakeGoHighLevelPabblyAI SDKRAGVisionVoiceSupabasePrompt Engineering
Method
- 01Measure the real cost of a process before adding AI.
- 02Aim for measurable time savings, not a wow-effect demo.
- 03Train the teams so they can maintain the tool after I leave.
- 04Prefer a system that abstains for lack of evidence over one that invents.
Agents & tooling
Agents that do the work, not throwaway prompts.
I apply AI to engineering itself: agentic workflows that run the repetitive, risky refactors at scale, without regressions, designed to be reused and spread across teams.
An agent is not a prompt: it is a tooled process — per-stage context, instructions and tools, review checkpoints, and metrics on real cases to know whether it delivers.
Real cases
Agentic workflows: analyze → plan → implement → verify
Per-stage context, instructions and tools · review checkpoints · measured on real cases
- Workflows structured into four explicit stages, each with its own context, instructions and tools, instead of one monolithic prompt.
- Checkpoints in between to review output, correct the trajectory and approve continuation — the agent never drifts unsupervised.
- Tests on real cases and metric collection to assess results, surface limits and steer the next iterations.
Stack Claude Code · AGENTS.md · specialized agents · slash-commands
Dependency-upgrade agent in CI/CD
Agent wired into the delivery pipeline · human validation after LLM analysis
- Dependency-upgrade agent wired straight into the CI/CD pipeline: the LLM analyzes and proposes, a human validates before merge.
- Designed from the start to be reusable internally, not for a single team.
Stack CI/CD · Claude Code · LLM
AI-assisted security review
Experiment · traceable reports
- Experiment on an AI-assisted security review pipeline, producing traceable reports.
Stack Claude Code · LLM
React 18 front-end migration
352 components · 4 simultaneous migrations · lead (65% of commits) · ~1 month
- Tech lead on a major migration run simultaneously (React 16→18, dead UI lib → mdb-react-ui-kit v10, Bootstrap 4→5, Enzyme→React Testing Library) over ~26,000 lines / 352 components, 250 files reworked.
- AI-assisted workflow (AGENTS.md pilot file + 4 pattern guides) ensuring regression-free code at scale, spread to other teams.
- ~−28,000 net lines of dependencies removed.
Stack Claude Code · Codex · Cursor
11 component families · ~7,800 lines of debt removed · 4 weeks
- AI engineering workflow built from scratch (2 autonomous agent sets + 4 slash-commands, assistant-agnostic) driving the full redesign of a React/TS app design system (~14,500 lines, 189 files).
- Ultra-incremental migration, one file at a time, until the old foundation was fully unplugged.
- Reusable toolkit (~1,900 lines of docs, 12 lessons-learned) spread to other teams.
Stack Claude Code · React 18 · TypeScript · react-dsfr
AI engineering stack
Claude CodeAGENTS.mdBMADSpecialized agentsSlash-commandsGitLab CI/CDDockerEvals & metricsCodexCursor
Method
- 01Industrialize AI engineering: reusable workflows spread across teams, not throwaway prompts.
- 02Break work into explicit stages, each with its own context, instructions and tools.
- 03Insert checkpoints: review, correct the trajectory, approve continuation.
- 04Measure agents on real cases: metrics, identified limits, iterations driven by numbers rather than intuition.