I design and ship production AI systems: agent orchestration, retrieval, and evaluation that hold up under real-world constraints. Hands-on across architecture and delivery, from first prototype to production.
I've built and scaled production systems across enterprise, growth-stage, and regulated environments, including a ~60-person engineering org at PwC, a Series A team I grew from 3 to 20+ at Spekit, and regulated platforms in fintech and healthcare.
Most recently I founded Beesla, an AI-native platform integrating agent workflows, retrieval pipelines, and production guardrails from day one, reaching 1M+ impressions across search engines and AI assistants and 80K+ unique users.
I work at the intersection of product decisions and engineering execution: translating business priorities into systems that ship cleanly, and sitting with the customers who use them.
Every fifteen minutes, Mender wakes up, reads the last hour of traces from another agent via the Arize Phoenix MCP server, clusters failures, hypothesizes a root cause, generates a focused eval set, drafts a prompt patch, re-runs the same evals against the patched version, and, if it measurably improves, posts a structured incident card to Slack with one-click human approval. He also reads his own traces every cycle and tunes himself: how many evals to generate, what confidence threshold to use, when to ask for help.
End-to-end verified: detected a real regression in a target agent ("ambiguous source currency silently defaulted to USD"), generated 10 focused eval cases, lifted pass rate from 4/10 to 10/10 on the patched version, a +60% lift, ready for one-click Slack approval.
The pattern generalizes: an agent that reads production traces, clusters failures, generates targeted evals, and validates its own patches before a human ever sees them. It's the same observe/evaluate/improve loop that any production LLM system needs.
Stack: Google ADK (Python) · Gemini 3 on Vertex AI · Arize Phoenix + Phoenix MCP · Slack Block Kit · Cloud Run · Cloud Scheduler · Firestore.
Independent studio work through Strawberry Digital: designed, built, and shipped in-house.
Interactive tournament-poker trainer. 136 lesson hands street-by-street plus a graded 100-hand exam covering push/fold, ICM, and M-zone play.
Curated events and places discovery across 50+ US metros. Chosen, not crowdsourced.
Automated product-research and deals engine with live marketplace pricing and buying guides.
Cozy offline math game for kids: arithmetic through fractions, decimals, and shapes.
A speed reader with a point of view: one word at a time, at a pace authored for each piece, with no library, no import, and no slider. You tune in, you don't operate it.
Call-and-response vocal-pitch trainer built on the classic expanding-scale warm-up. Plays each phrase on piano, listens as you sing it back, and scores every note in cents.
Designing and deploying autonomous agent architectures with orchestration, memory, and reliability patterns.
Context engineering, structured outputs, evaluation frameworks, and retrieval-augmented generation at scale. Hands-on across Anthropic Claude, OpenAI, and Google Gemini.
Scalable system design, microservices, API-first platforms, and operational reliability at enterprise scale.
AWS, GCP, and Azure architecture. Infrastructure-as-code, CI/CD pipelines, and cloud-native operational patterns. Python, TypeScript, PostgreSQL, Docker, Kubernetes, Terraform.
Team building, delivery discipline, cross-functional alignment, and shipping culture in fast-paced environments.
Production readiness frameworks, guardrails, risk assessment, and compliance in regulated domains.
Production over prototype.
Reliability over novelty.
Clear architecture over complexity.
Measurable outcomes over experimentation theater.