Hi, I'm Thomas 👋

I ship production AI systems end to end, and own the customer relationship that makes them work. Based in Raleigh, NC.

I built Meridian solo: a production, agentic GraphRAG platform on Vercel and Supabase. For five years before and alongside it, I've worked in customer-facing technical roles, currently leading Customer Success. ISC2 Certified in Cybersecurity.

The thing I do well is the part most engineers skip: take a genuinely complex AI system, make it real and reliable in production, and explain it to the person across the table.

🌐 See it live

The homepage at growmeridian.com renders a live, interactive slice of Meridian's actual GraphRAG knowledge graph. That's the real product, not a marketing graphic. Click around the graph to see how the retrieval layer connects.

🧠 What's inside Meridian

Designed, built, and operated solo, in production:

• Agentic AI pipeline: multi-provider, with a plan-then-execute tool-use flow, tiered model routing for cost control, strict JSON schemas, and streaming generation.
• GraphRAG retrieval: knowledge-graph expansion across thousands of curated relationship edges, multi-path retrieval fusion, signal-weighted re-ranking, dynamic confidence scoring, and gap detection.
• Agent context and memory: a separately namespaced embedding store with hybrid retrieval and reranking, served to internal agents through a custom Model Context Protocol (MCP) server with fail-closed auth.
• Multi-agent dev harness: subagents for context isolation, MCP integration across the toolchain, automated evals, and CI policy guardrails. It's the system that lets one engineer ship at speed.
• Eval and A/B testing: sticky variant assignment, z-test significance, and a ground-truth eval harness measuring retrieval quality on every change, so AI behavior is validated before it ships.
• Prompt-injection defense: a layered guard combining input classification, secret-leak detection, and source validation against retrieval bundles.
• Interactive knowledge graph: a WebGL graph explorer (Three.js and Sigma.js over graphology) rendering 60fps across tens of thousands of edges, now the homepage hero.
• Integration surface: a public HMAC-signed webhook ingestion API (signature rotation, idempotency, replay protection, per-key rate limiting) that external scripts and devices push to, plus embeddable tools with a developer API-key system.
• Internal tooling: a Postgres-native work-items tracker that replaced a third-party issue tracker (full history migrated losslessly), plus autonomous monitoring and daily briefs orchestrated via pg_cron.
• Security and observability: column-level access control verified by attack simulation, row-level security across all tables, Vault-managed scoped secrets, a GDPR/CCPA-ready consent and provenance model, CI security scanning, Sentry, PostHog, and per-call LLM cost tracking.

🔭 Currently building

A real-time autonomous agent that perceives a live application from raw screen pixels (computer vision), reasons over a graph-based world model, and learns human-like behavior through reinforcement learning, graded against an adversarial discriminator that scores how human its actions look. Sandboxed and self-hosted, in active development.

🤝 The other half of the job

I've spent five years on the customer side of technical products and currently lead Customer Success across a portfolio that runs from small teams to government. So I'm as comfortable sitting across from a customer and turning a hard technical problem into plain language as I am writing the code behind it.

🧰 Stack

• Languages: TypeScript, Python, SQL, PowerShell, Bash, JavaScript
• Frameworks and frontend: Next.js, Node.js, React, Tailwind, Zustand
• Infrastructure: Vercel, Supabase, AWS, PostgreSQL, pgvector, Edge Functions (Deno), Docker, GitHub Actions, pg_cron
• AI/ML: Anthropic Claude, OpenAI, Model Context Protocol (MCP), GraphRAG, RAG, agentic tool use, hybrid search, embedding pipelines, eval and cost ops
• Visualization: WebGL knowledge-graph rendering (Three.js, Sigma.js, graphology)
• Quality and security: Playwright, pgTAP, Husky, RLS, Supabase Vault, Sentry, PostHog, CI security scanning

🔗 Connect

• LinkedIn: linkedin.com/in/thomaskawas
• YouTube: youtube.com/@thomaskawas (building-in-public walkthroughs)
• Meridian: growmeridian.com
• About: growmeridian.com/author/thomas-kawas