I build and ship AI agents in production — and get my leverage from writing the loops that run them, not the prompts.

📍 San Francisco · markmeleka.com · LinkedIn

Software engineer at Carta — the platform companies use to manage their cap tables, valuations, and employee equity. I joined in data science and machine learning in 2019, founded what became the Total Compensation product, and have since shipped across four of its product areas. Over the past year I've led the technical push to put AI agents into Carta's products — and into the hands of its largest customers. I work across products, teams, and the executive level: proposing and shipping new directions, planning and directing work across teams, and mentoring engineers.

My leverage comes from loop engineering — building the harnesses that decide what to prompt, when, and whether the output is good enough, so agents do the work. It's why I'm among Carta's top engineers by merged-PR volume.

• Pitched and built Carta's first AI agent that works across its products — spotted the opening when single-team efforts had stalled, and took a hackathon prototype to a company-backed, multi-team initiative. Built on internal Model Context Protocol (MCP) servers.
• Shipped production agent infrastructure — an agent service with full observability and CI/CD, an in-product AI chat experience, and a secure auth pattern for agent-to-service calls across services.
• Designed the bridge from Carta's web app to an external LLM assistant — personalization and handoff, so the assistant knows the user and resumes where they left off. Shipped in 3 weeks.
• Created and lead "Working with Agents" — 101/102 trainings for technical and non-technical staff, delivered org-wide at a company of hundreds of engineers; plus AI-enablement on-sites for Carta's largest customers.

I take LLM and agent systems from demo to dependable — and I've done the research to prove how. My Oxford MSc dissertation (Software Engineering, 2025) built and rigorously evaluated a retrieval-augmented assistant over health records: a 540-run factorial study with formal statistical testing, plus a failure-mode taxonomy (hallucination, missed context, contradiction) tying design choices to reliability. The headline result — reasoning, not retrieval, is the bottleneck on hard queries — and the methods for grounding, citations, and latency/cost control are exactly what I bring to production agents.

Carta spans the lifecycle of company ownership. I've shipped across four of its product areas — plus the HRIS data layer that feeds them all, and the AI agents (above) that cut across everything.

• Total Compensation (pay-and-equity benchmarking) — founding backend engineer; shipped core features (international market adjustments, role taxonomy, variable comp) now used by thousands of companies across millions of employee records. Led its migration off a legacy, PII-laden data model onto governed data — shadow-mode rollout and live monitoring, for zero regressions and zero customer impact.
• Stakeholder Management (a company's equity holders and their grants) — led an equally involved migration off a legacy data model with major PII exposure, hardening the data behind ownership records.
• Cap Table (the system of record for who owns what equity) — built a cross-team equity-refresh feature spanning the compensation and cap-table systems, shipped ahead of schedule.
• LLC (ownership and tax for LLC-structured companies) — product extensions including self-service cap-table calculations (replacing per-client hard-coding) and a nearly 4× gain in tax-document matching accuracy; founded the team's technical-integrity program and made its CI ~20% faster.
• HRIS (the integration layer that feeds every Carta product with HR data) — led an org-wide migration across six major versions of a key HR-provider API, shielding thousands of customers from breaking changes with zero downtime.

Applied ML and data. Data scientist at Kik (2019) — led anti-spam ML across tens of millions of users, cutting the spam-report rate ~70% while holding false positives down. Technical Solutions Specialist at Sortable (2017–2018) — built prototype features, fixed client-side integrations, and surfaced revenue opportunities through data analysis.

Build with: Python · Go · TypeScript · Java · Django · LLM agents · MCP · RAG · evals · distributed services · Docker / Kubernetes · CI/CD

Studied: MSc Software Engineering, Oxford · CS & Mathematics, Waterloo · Data & Economics MicroMasters, MITx · BBA, Brock