Skip to content
WIZZARD.WORK

Applied AI engineering

Auckland / Aotearoa36.8509° S

WizzardWork.

Agents, pipelines, and the plumbing between them.

Selected work

I am Nico, based in Auckland. I build AI systems for support workflows, knowledge retrieval and device telemetry, then keep them standing once real data arrives.

See the workQuery my resume via MCP
  • RAG over support tickets
  • MCP servers
  • IoT telemetry at fleet scale
  • Backpressure and replay on ingestion
  • Python · SQL · Dremio · Snowflake
  • Agent evaluation harnesses
  • AI generated media
  • Next best action for support staff
02

Proof in numbers

  • 5+ YRS

    building data pipelines and AI systems

    Across telemetry, retail refrigeration, and internal tooling.

  • 25+

    ingress and egress APIs in production

    Automated ingestion, real-time processing, backend alerting.

  • 400K USD

    contract supported by cloud architecture I built

    Data onboarding design that closed in 2024.

    • Spanish
    • English
    • Portuguese
    • German

    languages spoken at working level

03

What I build

  • Agent systems

    Assistants that pick the next action instead of guessing. Tool routing over MCP, scoped permissions, and evaluation runs before anything reaches a user.

    • MCP servers exposing internal tools to models
    • Next-best-action recommendation for support staff
    • Deterministic evals with recorded traces per release
    Stack
    • MCP
    • Python
    • MindStudio
    • N8N

  • Retrieval that cites

    RAG built so answers point back at a document, a ticket, or a row. Retrieval quality is measured, not assumed, and every response carries its sources.

    • Chunking and reranking tuned per corpus
    • Citation required on every generated answer
    • Freshness checks so stale documents stop ranking
    Stack
    • RAG
    • MongoDB
    • Dremio
    • Python

  • Telemetry pipelines

    Device data that arrives late, out of order, and occasionally wrong. The pipeline reconciles it, flags what broke, and keeps operations able to see the fleet.

    • Ingestion APIs with backpressure and replay
    • Real-time alerting on device and fleet thresholds
    • Warehouse models that analysts query without help
    Stack
    • Snowflake
    • SQL
    • AWS
    • Power BI
04

Agent orchestration

A run that fails once and recovers

A replay of one agent run on a fixed timeline. The planner selects three tools. The ticket API returns 503 on the first attempt, so the planner schedules a retry and the run finishes on the second pass. Nothing here is random: every replay shows the same failure at the same second.

runrun_7c41f0elapsed+12.482s

Pipeline order. A request enters the planner. The planner calls a retriever, a SQL tool and a ticket API at the same time. The ticket API fails on its first call and the planner schedules one retry. A verifier checks all three results before the answer is returned.

INPUTRequestokAGENTPlannerokTOOLRetrieverokTOOLSQL toolokTOOLTicket APIokCHECKVerifierokOUTPUTAnswerok
Node state
  • idle
  • running
  • ok
  • failed
  • retry
Run log18/18
  1. +0.046srequestaccepted
  2. +0.428srequestpayload validated
  3. +1.016splannercontext requested
  4. +1.674splannercontext loaded
  5. +2.351splanner3 tools selected
  6. +2.913sretrieverquery dispatched
  7. +2.977ssqlquery dispatched
  8. +3.042sticketsGET /tickets sent
  9. +4.496stickets503 from upstream
  10. +5.128sretriever8 chunks returned
  11. +5.374splannerretry 1 of 2 queued
  12. +5.641ssql214 rows
  13. +7.614sticketsGET /tickets sent
  14. +8.801stickets2 tickets returned
  15. +9.366sverifiersources cross checked
  16. +11.023sverifier3 of 3 claims grounded
  17. +11.590sanswerstreamed to client
  18. +12.482sanswerrun complete
  • 5.888s in tool calls
  • 4.305s from 503 to recovery
05

Selected work

  1. 01

    Mahi

    A portable Windows utility that restores a full project profile with one keypress.

    • Rust
    • Tauri
    • SQLite
    • Win32
    Last commit
    Unavailable
  2. 02

    Cooler Capture

    A turf-war game layered on Bluetooth cooler visit data so field reps play by scanning.

    • Next.js
    • MapLibre
    • TypeScript
    • Python
    Last commit
    Unavailable
  3. 03

    HR onboarding wizard

    Onboarding as a state machine, with a model only where judgement is needed.

    • TypeScript
    • Next.js
    • RAG
    • Postgres
    Last commit
    2mo ago
See all work
06

Moving image

Films imaginedby people, madepossible by machines.

wizzard.mov

The moving-image side of my practice runs at wizzard.mov: AI film, direction first, models second. Same hands, different output.

Discipline
AI film and direction
Tools
Blender, Higgsfield, After Effects
Based
Auckland, New Zealand
07

Stack

01Languages
  • Python in a listed project
  • SQL in a listed project
  • TypeScript in a listed project
  • Bash
02AI
  • RAG in a listed project
  • MCP in a listed project
  • MindStudio
  • Evals
  • Prompt tooling
03Data
  • Dremio in a listed project
  • Snowflake in a listed project
  • MongoDB
  • Power BI
  • dbt in a listed project
04Automation
  • N8N
  • Airflow
  • GitHub Actions
05Cloud
  • AWS
  • Azure
  • GCP
  • Vercel in a listed project

Key

Underlined entries appear in the stack of a project on the work index. The rest do not.

08

Agent endpoint

Hire me through your agent

This site runs a Model Context Protocol server. Point an agent at the endpoint below and it can read my experience, stack, projects and availability without a human in the loop.

Tools
  • get_profile
  • list_projects
  • get_experience
  • get_skills
  • check_availability
Query my resume via MCP
POST/api/mcp
$ curl -s https://wizzard.work/api/mcp \  -H 'content-type: application/json' \  -d '{    "jsonrpc": "2.0",    "id": 1,    "method": "tools/call",    "params": {      "name": "get_profile",      "arguments": {}    }  }'
200 OKget_profile
{  "jsonrpc": "2.0",  "id": 1,  "result": {    "structuredContent": {      "name": "Nicolás López Guerra",      "role": "AI Solutions Engineer",      "location": "Auckland, New Zealand",      "email": "nicolopezg@proton.me",      "available": true    }  }}
JSON-RPC 2.0 over HTTP POST05 tools
09

Contact

Get in touch

Fastest route is email. I read everything and reply to anything with a specific question in it.

Email
nicolopezg@proton.me
Agent endpoint
curl -s https://wizzard.work/api/mcp -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"get_profile"}}'

ReturnsIdentity, role, location, contact and current availability.

AvailabilityOpen to AI engineering roles and contract work

Auckland, New Zealand