I build AI that works in production, and I measure it.

AfterQuery

Software Engineer, AI Evaluation & Benchmark Engineering

A Y Combinator-backed AI research lab. Work on Projects Silver, Pluto and Fenrir across two tracks: AI Model Evaluation & Benchmark Engineering, and Full-Stack Software Engineering (Testing & Evaluation).

• Author SWE-bench-style coding benchmark tasks: real bugs grounded in a repository's base commit, with machine-verifiable test harnesses (null/oracle-gated) and reference solutions.
• Work on C++/Rust systems and memory-safety benchmarks (Project Fenrir).
• Review instructions and rubrics against strict evaluation criteria (leakage, ambiguity, verifier strength, reproducibility) to keep AI training data reliable.
• Document findings, failure modes and reproduction steps as structured, specific feedback.
• Delivered accepted benchmark tasks across multiple repositories, earning per-task and per-repository bonuses.

sequa

AI Engineer

Built an agentic funding-application assistant for small Ethiopian businesses.

• Applicant agent turns a spoken story, phone photos and a paper licence into a complete, structured funding application, in English, Amharic, German and more.
• Unverified fields go on a gap list (what's missing, what it needs, from whom) instead of being guessed; declarations are explained in the applicant's language and never ticked for her.
• Reviewer agent scores each application against the funder's weighted 100-point grid with per-criterion reasoning, runs the eligibility gate and exclusion factors, flags contradictions, and returns a ranked shortlist a reviewer can defend.
• Stack: LangChain, FastAPI, Next.js.

AfriDev (Upwork agency)

Backend & AI Engineer

Freelance contract work through AfriDev, delivering Aquinas: Scholastic AI Tutor for an offshore client (private repository).

• Designed and built an end-to-end RAG tutoring platform with LlamaIndex, grounded in Catholic scholastic philosophy.
• Next.js frontend on a modular Django REST Framework backend.
• Celery with a Redis broker for background processing, plus Redis caching for queries and responses.
• Evaluated answer quality with an LLM-as-judge pipeline before release.
• Dockerized all services for production and delivered under a strict client deadline.

Blih Marketing and Communication PLC

AI Automation Engineer

Built an AI tender intelligence system that replaced the team's manual tender research.

• n8n, PostgreSQL and a Node.js/TypeScript Playwright fetcher, all in Docker Compose.
• The fetcher logs in to the tender portal, reuses and renews its session, de-duplicates listings and enriches each tender with its full scope of work.
• An LLM qualifies each tender against Blih's two service lines behind a fail-open prefilter that never silently drops a plausible tender.
• Found about 16% of tenders were near-duplicates getting inconsistent verdicts; added de-duplication before the AI step so each group gets one call and one consistent decision.
• Qualified tenders become Google Calendar deadline events and a Gmail summary; every run is logged.