Eric Holt

Featured Projects

TraceAgent — Observable Agentic Research Platform

Python · FastAPI · Groq · Tavily · React · PostgreSQL · Docker · AWS

Built around a four-stage pipeline — planner, search, reflection, and synthesis — where the reflection stage evaluates result adequacy and conditionally loops back to search (up to 3 times) with refined queries before report generation. Every internal decision is persisted to PostgreSQL and streamed to the browser in real time via WebSocket.

Deployed to Hugging Face Spaces (Docker) and AWS EC2 (t3.micro) + RDS PostgreSQL (db.t4g.micro, VPC-only) · CloudWatch Logs via watchtower · 6 custom CloudWatch Metrics per run (RunsCompleted, PipelineDurationMs, ReflectionLoops, and more) feeding a live pipeline dashboard · Public activity feed — browse all runs with full attribution · Fork any run with a modified query, lineage tracked · Side-by-side run comparison · Multi-stage build bundles React/Vite frontend with FastAPI backend

Live Demo · Code · README

EvalBoard — LLM Evaluation & Prompt Testing Dashboard

Python · FastAPI · React · Groq · PostgreSQL · Neon · Docker · GitHub Actions

Define test suites with inputs, expected outputs, and scoring criteria — then run them against one or more Groq models simultaneously. Results are scored by an LLM-as-judge and stream live to the UI via SSE.

LLM-as-judge scores each result 1–10 with pass/fail, strengths, weaknesses, and reasoning · Parallel model execution for side-by-side comparison · Prompt variant management with auto-versioning · Pass rate and score trend charts · CI/CD via GitHub Actions → Hugging Face Spaces · 37 pytest tests

Code · README

DailySignal — Automated AI News Briefing Service

Python · FastMCP · GitHub Actions · Groq · PostgreSQL · Neon · Resend · React · Recharts · Docker · Hugging Face Spaces

Twice-daily AI news briefing service — fully automated from fetch to inbox with zero manual steps. GitHub Actions cron pulls from HackerNews, Product Hunt, and five RSS feeds, deduplicates via SHA-256 hashing, synthesizes a structured briefing via Groq, and delivers it to subscribers via Resend.

GitHub Actions cron scheduling — no always-on server required · Cross-source deduplication using SHA-256 content hashes · Per-subscriber unsubscribe tokens with automatic deactivation after 3 failed deliveries · PostgreSQL audit log per subscriber per run · FastMCP server mounted at /mcp — three tools (get_latest_briefing, get_run_history, get_subscriber_stats) let any MCP-compatible AI agent query live briefing data and delivery stats directly · 55 unit tests, 98% coverage · Deployed to Hugging Face Spaces with Neon serverless PostgreSQL

Code · README

SupportDesk — AI-Powered Customer Support Automation

Python · FastAPI · Groq · pgvector · React · Vite · Tailwind CSS · PostgreSQL · Docker · GitHub Actions

A webhook fires when a customer submits a ticket. The backend classifies it, searches a vector knowledge base for relevant documentation, and drafts a grounded reply — all before a human agent opens the ticket. The agent reviews the draft and cited sources inline, edits if needed, and approves. The reply goes out via Resend. Every pipeline stage is logged and broadcast to the live event dashboard in real time.

HMAC-SHA256 webhook verification with timestamp replay protection · Groq LLM classification (type, urgency, confidence) · fastembed + pgvector cosine similarity RAG retrieval · SSE streaming for draft tokens · WebSocket live event log · Resend email delivery · 71 pytest tests · Deployed to Hugging Face Spaces (Docker)

Live Demo · Code · README

AgentMesh — Multi-Agent Code Review & Repair System

Python · LangGraph · FastAPI · Groq · React · Tailwind CSS · Docker

Three specialized agents collaborate via a LangGraph StateGraph — Reviewer identifies issues, Fixer produces corrected code with a changelog, and Evaluator scores the fix across three dimensions and can reject the output and route back to Fixer with specific feedback for a second pass. Every agent decision and handoff streams to the browser in real time via SSE.

Evaluator scoring: Correctness (40%) · Completeness (40%) · Code Quality (20%) — score 50–69 triggers Fixer retry with evaluator feedback, capped at 2 passes · Structured Pydantic schemas for all inter-agent state · File upload or JSON input, language auto-detection · Stateless backend · Multi-stage Docker build deployed to Hugging Face Spaces

Live Demo · Code · README

Get Set Wedding — Production SaaS Planning Platform

React · TypeScript · Supabase · Stripe · Node.js

Built end-to-end as contract full-stack developer — database schema, PostgreSQL RLS for structural multi-tenant isolation, Stripe billing with webhook-driven server-side enforcement, and a context-aware AI assistant integrated with live user planning data.

PostgreSQL-layer RLS (structural isolation, not application-level) · Webhook-driven Stripe pipeline — server-side enforcement on every API request prevents client-side bypass · Refresh-safe JWT session rehydration · Demo-to-account conversion with isolated storage scopes · RBAC admin routing · AI assistant with live access to user planning data (budget, guest list, vendors)

README

GitHub Activity

About

I'm an AI Engineer at MarineMax, where I led a from-scratch migration of the company's production multi-agent sales assistant off a managed orchestration platform onto a LangGraph pipeline calling Claude directly — building full feature parity with claude.ai (Projects, Artifacts, in-place document editing, prompt caching) and running a self-directed hardening pass that closed eight concurrency and security issues before they reached customers.

Outside of work, I design and build full-stack AI systems with a focus on real-world deployment, observability, and production reliability. My personal projects span multi-step agentic pipelines with conditional looping and self-evaluation, RAG systems with pgvector and cited retrieval, React/Vite frontends with live SSE and WebSocket streaming, and automated scheduled pipelines with multi-source aggregation and email delivery. One project exposes a live MCP server via FastMCP, making pipeline data queryable by any MCP-compatible AI client. On the infrastructure side I instrument systems with custom CloudWatch metrics and live dashboards, ship with CI/CD, and have deployed to AWS. I also have production SaaS experience through Get Set Wedding, where I built Stripe billing enforcement, PostgreSQL RLS architecture, and a context-aware AI assistant against live user data. Everything here is production-grade with monitoring, tested codebases, and real users - not just demos.

Certifications

• · Model Context Protocol: Advanced Topics — Anthropic
• · Introduction to Model Context Protocol — Anthropic
• · Claude with the Anthropic API — Anthropic
• · Python Certified

Contact

Email · GitHub · LinkedIn