AI Engineer
Hire an AI engineer who ships to production
Not a prompt consultant. An engineer who puts an LLM inside a process that already runs money through it — with cost routing, grounding, guardrails, escalation to humans, and a test suite that catches regressions before customers do.
The part most people get wrong
The demo always works. What breaks in month two is cost (an unrouted Sonnet-only pipeline is 4x more expensive than it needs to be), hallucinated prices that turn into refunds, and the absence of any way to know the AI gave a bad answer until a customer complains.
What you get
- →Complexity-classifier model routing — Claude Haiku 4.5 for simple turns, Sonnet 5 for reasoning. Measured 74% token cost reduction over 4 weeks of production traffic with no measurable quality drop.
- →RAG grounded on the live catalog, with the assistant instructed to refuse rather than guess when the source cannot confirm a SKU or a price.
- →Hard guards against hallucinated prices, stock levels and promises the business cannot keep.
- →Multi-channel unified inbox — WhatsApp Cloud API official, MercadoLibre, TiendaNube, Facebook, Instagram, web widget.
- →Human handoff that fires before the customer gets frustrated, with full conversation state pre-loaded for the agent.
- →Audit log of every AI message and every prompt version, so a bad answer on Tuesday is traceable to the prompt that caused it.
- →263-test Playwright E2E sweep before every deploy.
- →The Anthropic API key lives in your account, billed to your card. No markup, no lock-in.
Proof — shipped and in production
Servifibras CRM + AI
A composites shop was renting an AI assistant. Their own Claude-powered agent now handles sales across 5 channels, with hard guards against price hallucination.
Virtual Commercial Assistant SaaS
SMBs (delivery, barbershops, clinics, retail) needed a WhatsApp assistant with STT and per-client JSON config. MVP built for stability over intelligence.
Educational Chatbot
A single-lesson pilot needed to prove the interaction model before scaling. Multiple-choice mechanics, lesson import system and lightweight CMS designed.
LINCE Water Monitoring
Water treatment failures were found after the damage, through paper reports. AI now surfaces anomalies from live data while they are still cheap to fix.
Frequently asked
Which model do you build on?+
Anthropic Claude Sonnet 5 for reasoning and Haiku 4.5 for routing and simple turns, selected per turn by a complexity classifier. GPT-5 and Gemini are wired as fallbacks when a client requires vendor diversity. For strict data-residency requirements, Llama 3.3 70B via Ollama on a dedicated GPU droplet.
How do you stop the AI from inventing prices?+
The assistant never has a price list in its context window. Every price question triggers a RAG lookup against the live catalog, and the system prompt instructs it to refuse and escalate if the catalog does not confirm the SKU. This single design decision is the difference between a sales AI and a liability.
What does a production AI system actually cost to run?+
The Servifibras CRM replaced a USD 2,000/month AI SaaS and runs at roughly USD 180/month in Anthropic API spend at ~74% Haiku routing. Exact numbers depend on conversation volume, but routing is the lever that moves it most.
This role is also called
- AI expert
- AI professional
- machine learning engineer
- LLM engineer
- chatbot developer
- AI automation specialist
Start with a written brief
Describe what the business is losing today and what the target outcome looks like. I come back with a concrete plan against that document — in English, Spanish, Portuguese or Polish.
Other roles
Full-Stack Developer
One person who owns the whole thing — database schema, API, frontend, deployment, and the production incident at 2am. No handoffs between a backend contractor and a frontend contractor who blame each other when something breaks.
Blockchain Developer
Contracts that shipped to mainnet and held real value — not a tutorial fork with the variable names changed. Vetra, a USD-backed stablecoin on Polygon with Chainlink Functions verifying reserves in real time, is the reference.
Security Engineer
When a production server is owned, the clock matters more than the credential. I do the full incident response workflow — evidence-preserving triage, eradication, credential rotation, patching (including hand-written patches when the vendor fix has not shipped), and hardening that makes round two expensive.
Business Automation Engineer
Not a Zapier consultant. An engineer who replaces the spreadsheet, the WhatsApp group and the clipboard with software the team stops having to think about — written in code you own, on infrastructure you control, with no per-execution billing.
Fintech Developer
The money layer, built with the paranoia it deserves. Signed webhooks, idempotent handlers, integer minor units, append-only ledgers, row-level locks. The boring rules that prevent the entire class of "the numbers do not add up" tickets.
WhatsApp API Developer
Official WhatsApp Business Cloud API via a Meta-approved BSP. Not Baileys, not Evolution, not a QR-scanning wrapper that gets the client's number banned in a week. The integration that survives contact with a real business.
Industrial Automation & IoT Engineer
Sensors are cheap now. The expensive part is turning a telemetry firehose into an alert that reaches the right person before the damage is done — and not one alert per reading, which trains everyone to ignore them.
Booking System Developer
The booking engine your business actually runs on — owned outright, not rented per seat. Real-time calendar sync, WhatsApp reminders in your voice, payments into your own account, and a cancellation queue that refills the slot before the front desk even notices.