AI Consulting · Claude · US & Latam

I take Claude to
production.

No demos. No slide decks. Real systems in real businesses across the United States, Colombia and Mexico.
I'm Carlos Betancur Gálvez, founder of btodigital and one of the leading specialists in implementing AI with Claude (Anthropic).

8
Proprietary AI products in production
11
Countries served across Latam and the US
15+
Years in digital marketing
$3k
Projects from USD for small businesses
What is this

What is an AI consultant with Claude?

An AI consultant with Claude is a specialist who designs, builds and ships production systems on top of Anthropic's Claude models, rather than selling generic "AI strategy." In practice that means WhatsApp and chat agents, RAG systems over your own documents and CRM, process automation, and CRM/ERP integrations, engineered on cloud infrastructure with real observability, security and predictable costs, not just a demo.

Why Claude

Why Claude and not ChatGPT, Gemini or an open model?

Model choice isn't ideological. It's strategic. For the enterprise use cases I see most, Claude offers three concrete advantages that end up making the difference in production.

Capability
Claude (Anthropic)
ChatGPT (OpenAI)
Gemini (Google)
Context window
Up to 200k tokens (full manuals, contracts, historical data)
128k tokens on GPT-4o
1M tokens (leader, but pricier)
Neutral Latam Spanish
Excellent; doesn't force Spain-flavored Spanish
Good, but occasionally shifts tone
Good, integrated with Vertex AI
Agentic reasoning (multi-step)
Leader at following long instructions and tool use
Solid, especially with o1
Improving fast, still behind on agents
Cost at typical enterprise usage
Medium-high, justified by quality
Variable by model
Cheaper on Flash, perfect for batch
Data privacy by default
Doesn't train on your API calls
Doesn't train on enterprise API
Encryption and IAM controls on GCP

In practice, my architectures are usually hybrid: Claude for primary reasoning and conversation, Gemini Flash for cheap high-volume batch processing, and open models on GCP when extreme privacy or minimal latency is required. The question isn't "which model is best?" but "which combination solves this specific problem at the lowest possible cost?".

In production

What I've built (and is running right now).

I combine one public case that lives on this very domain with five enterprise projects under NDA, presented by industry with technical detail and outcome.

Public case · My own agent

Beta — WhatsApp agent with Claude

What it does: greets anyone who messages my WhatsApp, replies in real time, identifies the lead's intent, qualifies them as A, B or C, sends PDF books, follows up when a conversation goes quiet, and escalates to me by email when someone wants to book a meeting.

Technical stack: Claude Sonnet 4.6 + Gemini 2.5 Flash + Firestore on Cloud Run, integrated with Kapso for the WhatsApp Cloud API. RAG over 500+ proprietary documents, audio support, a three-stage follow-up system (30 minutes, 3 hours, 6 hours), and quiet-hours respect (no messages between 10 p.m. and 7 a.m. Colombia time).

Result: operating cost between USD 5 and 8 per month. It captures, qualifies and routes leads 24/7 with no human intervention. Try it by messaging on WhatsApp →

Healthcare · Vertical platform

AI-powered doctor search engine

A vertical healthcare platform with 400+ published physicians. I built a system that automatically translates Google My Business reviews into English with Claude, generates personalized landing pages for VIP doctors with i18n, and monitors Google Ads campaigns daily with Cloud Run Jobs and smart alerts that notify when performance drops.

Stack: Express + Tailwind + Firestore + Cloud Run on GCP, Resend for transactional email, Vertex AI for batch processing.

Fashion retail · In-store QA

In-store sales quality SaaS

A multi-tenant platform that listens to recordings of in-store salesperson conversations and measures service quality with AI. It detects whether they greeted the customer, asked the right questions, handled objections, and closed cross-sell opportunities. The first client operates three locations in Medellín; the product is designed to scale to any retailer with in-person sales staff.

Stack: OpenPlaud fork + Vertex Gemini 2.5 Flash + Firestore + Google Login authentication. Processes hours of audio daily in batch.

Heavy industry · Commercial intelligence

AI data pipeline

Two-way sync between Bitrix24 and BigQuery every 12 hours for a leading industrial company in Latam. Executive dashboard in Next.js + Firebase with AI analysis over the sales funnel. We process around 50,000 historical deals to identify closing patterns, abandoned leads and churn.

Stack: Cloud Run Job + Cloud Scheduler for ETL, Firestore as operational cache, Vertex AI for analysis, Next.js + Firebase for the frontend.

E-commerce · Systems integration

Shopify ↔ corporate ERP middleware

An electronics distributor with a large catalog needed to sync Shopify with Microsoft Dynamics 365. I built middleware with paid-order webhooks, real-time shipment tracking updates, dual-SKU handling, and automatic fallback to Pub/Sub queues with exponential retry when the ERP doesn't respond.

Stack: Node.js on Cloud Run us-central1, Pub/Sub, Firestore, integration with Shopify and Dynamics APIs.

Professional services · Internal assistant

Productivity assistant with RAG

With my guidance, a services company built an internal assistant that answers questions about procedures, active contracts, internal policies and past cases. RAG over 2,000+ documents. It reduces the time the team spends hunting for scattered information and frees up the back-office.

Stack: Claude + Vertex embeddings + Firestore Vector Search + web interface in Astro.

Marketing · RAG over CRM data

Real-time marketing intelligence over your CRM

I've built RAG systems connected to platforms like HubSpot, Clientify, Klaviyo, Shopify and WordPress so marketing and sales teams can query their own data in plain language: customer history, audience segments, campaigns sent, purchase behavior. The team asks and the system answers with real data, not generic model generalizations.

Stack: Claude + Vertex AI embeddings + Firestore Vector Search + incremental sync via webhooks and each platform's REST APIs. The index updates automatically as CRM records change.

Public portfolio · btodigital

Nine products in production you can see and use today.

I don't just advise: my team at btodigital and I design, build and maintain AI SaaS products. Some are paid; six are completely free so you can see the quality before you talk to me.

SaaS products

Free AI tools

Every tool is fully functional, with no mandatory sign-up and no credit card. If you like them, let's talk about how to build something like this inside your company.

Use cases by industry

Where Claude fits in your sector.

Not every use case fits every industry. These are the patterns that work best in each vertical, based on projects I've executed or advised on over the past few years.

🏥 Healthcare

  • RAG over medical records and clinical protocols for diagnostic support
  • WhatsApp agents to book appointments and answer common patient questions
  • Call-quality analysis in medical contact centers
  • Translation and localization of content for international patients

🛒 Retail & E-commerce

  • Bulk optimization of product listings with AI (ShopiUP)
  • Automated post-sale support with escalation to humans
  • Personalized recommendations based on real behavior
  • In-store sales QA with audio analysis

💼 Professional services

  • Internal assistants with RAG over contracts, policies and past cases
  • Automatic generation of proposals, memos and reports
  • Analysis of RFPs and lengthy documents
  • Lead qualification in forms and conversational agents

🏗️ Industry & manufacturing

  • Commercial intelligence pipeline with CRM and BigQuery
  • Analysis of public and private tenders
  • Detection of churn, leakage and closing patterns in the funnel
  • Technical assistants for field teams with full manuals in RAG

🎓 Education

  • AI grading systems with instant feedback
  • Personalized tutoring that scales to thousands of students
  • Generation of pedagogical content tailored to levels
  • Learning analytics and dropout prediction

🏦 Financial services

  • Automated document analysis for risk and compliance
  • Fraud detection based on conversational patterns
  • Assistants for advisors handling complex products
  • Processing of customer inquiries with smart classification and routing
How I work

A four-phase process. No empty promises.

I don't promise magic or 24-hour transformations. A well-executed AI project has clear phases, measurable deliverables, and the ability to stop the process if reality doesn't match the expectation.

01

Diagnosis

1 week

An in-depth interview with you and your team. I identify the real problem (not the symptom), assess whether AI is the right solution or whether there's a simpler fix, and deliver a document with three possible paths: ambitious, balanced and conservative. If you don't want to move forward after this, no problem: the diagnosis is yours to keep.

02

Proof of concept (POC)

2 to 3 weeks

I build a functional but scoped version that proves the solution works in your real context. We test it with a pilot group, measuring quality, cost and adoption. Here we kill bad ideas quickly instead of spending months building something nobody will use.

03

Production implementation

4 to 6 weeks

We take the POC to a robust system: architecture on Google Cloud Platform, observability with Cloud Logging, an alerting system, access controls, error handling and predictable costs. I document everything so your team can maintain or evolve it without depending 100% on me.

04

Continuous iteration

Monthly or quarterly

AI isn't a project that ends: it's a system you tune. We refine prompts, update the RAG, add capabilities, monitor metrics and adjust costs. This can be a monthly retainer or one-off sessions as needed. You set the pace.

Transparent pricing

What it costs. No fine print.

This is one of the questions I most dislike dodging. Here are the real ranges for my projects, in US dollars, so you know whether it makes sense to keep talking.

Initial diagnosis
Free

A 30-minute session over video call or WhatsApp. I help you determine whether your use case makes sense for AI, what order of magnitude it might cost, and what returns are realistic.

Proof of concept
USD 3,000 to 8,000

A functional but scoped version. Ideal for small businesses with a well-defined use case, or for corporations that want to validate before committing a large budget. Timeline: two to three weeks.

Full implementation
USD 10,000 to 30,000

A production system with robust architecture, documentation, observability and a maintenance plan. The vast majority of enterprise projects fit within this range.

Monthly retainer
From USD 1,000/mo

Ongoing maintenance, iteration, evolution and support. Variable depending on scope, SLA and the level of involvement from my team.

Compared with Anthropic's Forward Deployment model

Anthropic offers its own implementation service (Forward Deployed Engineers) directly out of San Francisco, with rates that typically start at USD 150,000 per project and focus on Fortune 500 companies. My value proposition is to give you the same technical level, applied to your market, at prices 10 to 50 times lower, in fluent English and neutral Spanish, and with an understanding of the local context.

Frequently asked questions

What I get asked most.

What do I need before implementing Claude in my business?

The only essential is a clear use case: a repetitive process, an operational bottleneck, or a conversation you are losing. You don't need perfect data or an in-house technical team; I help you start from scratch. What does help is decision-making at the leadership level: implementing AI well changes processes, it doesn't just add another tool.

How soon do I see results with Claude?

A proof of concept (POC) takes two to three weeks. A production-ready implementation takes four to six weeks. In most of the projects I've worked on, the client sees the first signs of savings or commercial impact within the first month. Fine iteration (tuning the prompt, refining the RAG, adding capabilities) is continuous.

Do you work solo or do you have a team?

I personally lead every project, but behind me is the technical team at btodigital, my agency. Depending on the scope, I bring in developers, designers, and data or paid-media specialists. I always build the strategy and architecture myself: I don't delegate the piece that decides whether the project lives or dies.

Is my data safe with Claude?

Anthropic, the company behind Claude, does not use your API call data to train its models by default. On top of that, in enterprise projects we architect the solutions on Google Cloud Platform with encryption at rest and in transit, IAM-based access controls, and strict tenant separation where applicable. Every project includes its own data retention and deletion policy.

Why Claude and not ChatGPT or OpenAI?

Claude has three concrete advantages for businesses: a huge context window (it processes entire long documents without losing the thread), excellent performance across English and neutral Latin American Spanish, and more mature agentic capabilities for complex workflows. In projects where I've compared Claude with GPT-4 and Gemini, Claude wins on extended reasoning and on following multi-step instructions.

Do you serve businesses in the United States?

Yes. I handle projects in English and have clients in Miami and other U.S. cities. My site is available in English and Spanish, and both Beta (my WhatsApp agent) and the solutions I build support both languages natively. Projects are managed remotely.

Do you serve small businesses or only large corporations?

I work with both. My pricing model is designed so that a small business with a well-scoped project can have AI in production for under USD 10,000, while I also run corporate projects worth several tens of thousands of dollars. The difference is not the size of the client but the depth of the problem and the complexity of the integration.

Want to see whether it makes sense to implement Claude in your business?

Three paths depending on where you are. I answer personally.