AI products
LLM features people pay for: prompt design, structured output, evals in CI, cost tracking per call.
AI engineer, Philippines
Senior full-stack engineer building LLM products, agents and business automation end to end: from the prompt and the evals to the deploy and the bill.
1,000+ commits on Kaivi93% to 100% recall@1 with a reranker, first 15-question RAG eval703 unit tests on GroundworkClaude Code native workflow
Selected work
Six systems, end to end
Each one has a write-up: the problem, the architecture, how the AI part is built and measured, and what I would change.
- 01
Kaivi
AI mock interviews from your resume and the job description, with a coach's report.
- 02
QuickSite
Pipeline that builds and reviews demo websites for local businesses, then runs the outreach.
- 03
AI Engineering Lab
Hybrid RAG with reranking and an LLM-as-judge eval harness, plus LangGraph agents and MCP.
- 04
Groundwork
Turns a messy idea into phases, tasks and risks as a reviewable diff, then hands off to Claude Code.
- 05
Flightdeck
Control tower for every Claude Code session on a machine, with a live terminal in the browser.
- 06
Paperweight
Edit real PDF text in the browser with PDFium WASM and on-device OCR. Nothing is uploaded.
What I do
Three things, done properly
Business automation
Pipelines and agents that take repetitive work off a team: generation, review, outreach, reporting.
AI-native workflow
I build the tools I work with: Claude Code orchestration, hooks, planning agents and coaching.
How I work
Agents do the typing, I own the result
Every project above was built with a Claude Code workflow I designed: specs before code, parallel agents, reviewers and gates.
Spec first
A written design gets approved before any code. Decisions are recorded with the reason.
Parallel implementers
Independent tasks run as separate agents on separate files, then one integration step.
Evals and hooks as gates
Prompt changes run evals in CI. Hooks block destructive commands and run verify on stop.
Measured, not vibes
Token cost, latency and quality are tracked, so model choices are defended with numbers.
More work
Tool
Claude Coach
Reads real Claude Code prompts and sessions, compares them with official guidance, and plans fixes with rollback.
Client projectBizDesk
Branch sales counter and database-enforced access, built for an owner who cannot be at every branch.
Client project
npm packageCoaching team site
Bilingual English and Arabic marketing site with right-to-left layout, built static on Astro 7.
claude-adapt
Scores a codebase and generates a tuned CLAUDE.md and skill packs for it.
Contact
Let's buildthe next one
Hiring
Senior AI engineer roles
Remote or hybrid. Happy to walk through any of these systems in an interview.
Projects
Contract work
AI features, automation and agent workflows for teams that need it shipped, not just prototyped.