Services
AI that does something useful in production.
I build LLM-powered apps, autonomous agents, and AI tooling that runs in the real world. Not demos. Not experiments. Systems your team actually uses.
Tell me what you're trying to buildThere's a lot of AI work that looks impressive in a video and falls apart in production. I've shipped enough of it to know where the gaps are.
What I build
LLM-powered applications
Chat interfaces, document Q&A, natural-language over your own data. If you want your users or your team to talk to your database, your docs, or your product in plain English, that's the core of what I build.
Autonomous agents and multi-agent systems
Agents that reason over a task, call tools, handle retries, and don't require a human in the loop for every step. I've shipped a six-agent system running on LangChain and LangGraph with CrewAI orchestration, coordinated via Temporal.io durable workflows — when one step fails, the whole thing doesn't fall over.
MCP servers
Model Context Protocol servers that give Claude or any compatible model live, structured access to your databases, APIs, or internal systems. I've built four production MCP servers across Python and Node — including one running Pinecone vector search over a 352K-domain index.
Voice AI
Real-time voice agents bridging LLMs to live phone systems over SIP/VoIP. I built a Gemini Live–to–SIP bridge with real-time audio resampling for a production phone system — if you need an AI that answers, routes, or gathers information over a call, I know what the latency and resampling problems look like before you hit them.
AI consulting and architecture
Not sure what you actually need? That's a valid starting point. I can audit what you've already built, help you decide between off-the-shelf tools and custom code, and tell you where the real complexity is hiding before you commit to a build.
Workflow automation
n8n, API chains, BullMQ queues, SSE for live progress. Repetitive processes that currently require a person sitting in the middle — if it follows a pattern, it can probably run on its own.
Some of what I've shipped
Six coordinated agents running on LangChain and LangGraph with CrewAI orchestration. Multi-model: Claude, GPT-4, and Gemini. Visual pipeline builder in ReactFlow with a Monaco editor for custom logic. Live integrations with QVC, Tractor Supply, Cin7, ShipStation, and QuickBooks. Durable workflows via Temporal.io, so nothing gets lost when a step fails.
Internal chat interface over 106K+ client records for a US digital-marketing agency. Ops team asked questions in plain English, got streaming answers with auto-rendered charts. Built on Next.js 15 with a 70-component shadcn/ui library. Used in production daily.
Four production MCP servers across Python and Node, giving agents structured access to live databases, AI-generated SQL, and a 352K-domain Pinecone vector index — enabling autonomous data retrieval without a human running queries by hand.
Production voice agent
Real-time voice agent bridging Gemini Live to a production phone system over SIP/VoIP, with real-time audio resampling to keep the call responsive. Live in production — the same architecture applies to any similar voice-AI build.
Connect any Postgres or MySQL database, ask questions in plain English, and Claude generates read-only SQL — results auto-chart in the interface. AES-256-GCM encryption on stored credentials, with a SQL safety layer blocking writes. Public on GitHub: github.com/m4k00.
Stack
LLM providers
- OpenAI (GPT-4, embeddings)
- Anthropic (Claude)
- Google (Gemini)
- multi-model setups
Frameworks and tooling
- LangChain
- LangGraph
- CrewAI
- MCP (Model Context Protocol)
- Vercel AI SDK
Vector and memory
- Pinecone
- pgvector
- Redis
Voice and real-time
- SIP/VoIP
- real-time audio resampling
- streaming responses
Orchestration and queues
- Temporal.io
- BullMQ
- n8n
- Pub/Sub
Backend infrastructure
- FastAPI
- Node.js
- PostgreSQL
- Supabase
- GCP (Cloud Run, BigQuery)
- Docker
Not sure what to build?
Most people who reach out already have one of three situations:
- They have an idea but don't know whether to use an existing tool or build something custom. Usually I can answer that in a 30-minute call.
- They have a prototype that works in a demo but breaks under real conditions. This is a solvable problem, and it's usually one or two architectural decisions that caused it.
- They have a working system but want to add AI on top of it without rebuilding everything. This is often more doable than it sounds, and less doable than vendors will tell you.
I don't have a retainer model for consulting that goes on indefinitely. I help you figure out the build, scope the real work, and either do it or hand you off with a clear plan.
Background
Microsoft Certified: Azure AI Engineer Associate. 20 years in production as a full-stack engineer before most of today's AI tooling existed — which means I build AI features on top of real production infrastructure, not the other way around. I've shipped four production MCP servers, a six-agent system running in production, and a real-time voice agent bridging an LLM to a live phone system.
Based in Tamaulipas, Mexico. Bilingual English and Spanish. Working US hours.
FAQ
Do you build AI agents, or AI features inside an existing product?
Both — most of the time they're the same project. Multi-agent systems get built inside a real product, with the data model and API layer to support them, not as a standalone demo.
Which LLM should I use — GPT, Claude, or Gemini?
Whichever fits the task and the budget. I've shipped production systems on Anthropic, OpenAI, and Gemini, sometimes routing between them in the same platform. There's no vendor lock-in in how I build.
What is an MCP server, and do I need one?
An MCP server is a structured interface that gives an AI agent live access to a database, API, or tool — instead of the agent guessing or working from a stale snapshot. You need one if an agent has to act on your real, current data.
Is this just wrapping ChatGPT in a chat widget?
No. A chat widget is the easy 10%. The other 90% is orchestration, retries, grounding the model in your actual data, and making sure a failed step doesn't take down the whole workflow — that's the part I build.
Do you do AI consulting without writing code?
Yes. Architecture reviews and build-vs-buy assessments are a standalone engagement — a written recommendation on what to build, what to buy, and where the real complexity is hiding, with or without a build attached.
Do you do proof-of-concepts?
I build thin vertical slices instead: one real flow, in production, behind a flag. It answers the same question a PoC would, without becoming throwaway code.
Who maintains the system after launch?
I do, under retainer if you want ongoing support. The systems above run in production today — infrastructure included.
Is there a minimum project size?
AI work generally starts around $5,000 — smaller builds are usually better served by an off-the-shelf tool, and I'll say so if that's the case. Send it anyway; the budget-range question on the contact form helps me answer honestly and fast.
Need the full infrastructure underneath this? That's the Full-Stack Development side.
Tell me what you're trying to build.
Whether you have a full spec or just a problem you want to solve, I can tell you what's realistic, what it'll take, and whether I'm the right fit.
I reply within one business day.