AI SYSTEMS ENGINEER
Hire an engineer who ships working AI systems.
I build the agent, the retrieval layer behind it, and the production application around it. Four of these are running on real domains with public repositories, so you can check the work before we speak. Retrieval is benchmarked, usage and latency are reported, and the code and the deployment are yours. Fixed-price, remote worldwide.
See it running on your own site first
Paste a website and I'll build a live agent trained on its public pages in front of you. It is the same crawl and retrieval path the production systems use, not a scripted recording.
What I get hired to build
Retrieval and document intelligence
Extraction, sentence-aware chunking, embeddings and vector search with an HNSW index. Answers arrive with the source attached, and retrieval is checked against a goldset rather than eyeballed.
Tool-using AI agents
LangGraph state machines with real tool-calling loops, conditional routing and validated tool arguments. Handoff to a human carries the full conversation.
Receptionists and support agents
Always-on agents for small and mid-size businesses. Answer, book, capture, route and escalate, embedded on your site with one line of script.
MCP server development
Model Context Protocol servers that put your systems behind a typed tool interface, so AI assistants can read your data and run your workflows directly.
AI added to an existing stack
Connect models to the database, API and CRM you already run. No rebuild of what works, no new platform subscription.
Internal tools replacing manual work
Approval queues, intake forms and reporting on your own data, with auth, roles and an audit trail. Supporting role: the web and mobile app around the AI.
How it works
Tell me what you're building
Share the brief. What it must do, what data it reads, what it has to be able to explain.
I review the scope
Feasibility, architecture, and a fixed-price proposal with clear milestones. If AI is the wrong answer here, I say so at this step.
We build in milestones
Working software at each milestone. Retrieval gets tested against a set before it ships, and token and latency reporting goes in early.
You get the production system
Code, deployment, documentation and reporting, all handed over. The system and the infrastructure account are yours.
Proof of work, live in production
Each of these is a running deployment with a public repository, not a prototype. Open any of them and use it.
AI Customer Support Agent
A LangGraph agent with a real tool-calling loop and conditional routing. It answers from a vector knowledge base, checks orders, creates tickets, and escalates to a human with the full conversation attached. Embeddable with one line of script.
Smart PDF Workspace
Document intelligence with sentence-aware chunking, OpenAI text-embedding-3-small embeddings, and pgvector cosine search over an HNSW index. Every answer carries its source. Retrieval is benchmarked separately on the AI Virtual Receptionist knowledge base.
AI Virtual Receptionist
A 24/7 receptionist that answers from a vector-retrieval knowledge base, books appointments, captures and qualifies leads, routes visitors to the right department, and hands off to a human with an AI-drafted reply ready to send.
HireMe MCP Server
MCP server proof rather than agent work. Five tools, four reads and one rate-limited write, exposed over the Model Context Protocol so any compatible assistant can use them. It makes no model calls of its own.
How the claims get checked
Retrieval is benchmarked
A goldset of queries with known-correct sources, run offline through the pipeline, computing precision, recall and F1 with a threshold sweep. The benchmark runs on the AI Virtual Receptionist knowledge base (dental clinic KB, 15 chunks, 18 queries) with a deterministic offline embedder. Most recent vector retrieval run: 93.8% F1 at threshold 0.35. This covers retrieval policy only, not the production embedder (text-embedding-3-small) or written answer quality.
Cost, latency and steps are reported
Every model call records token counts and latency, persisted and exposed on an authenticated metrics endpoint, and every agent run emits a per-step trace: one span per graph node, per model call and per tool call. That is step-level visibility rather than full distributed tracing, and I describe it that way. Visitor content is redacted before export by default.
The code is public
All four systems have public repositories. Read them, fork them, or run them yourself.
Multi-tenant by default
Short-lived JWT access tokens with bcrypt hashing and row-level isolation between tenants on all four. The receptionist and document workspace go further, with rotating refresh tokens and server-side revocation through a token version.
Have a project in mind?
Tell me what you need built. I'll review the scope and send a fixed-price proposal with clear milestones.
Describe your project