Skip to content

Full Stack & AI Engineer · Lima, Perú

Joseph Robles

Forward Deployed Engineer (FDE)

Every project, role and skill here is a node. Scroll to walk the graph, click a node to open it.

Currently

  • Entrepreneur in Residence at Xertica.ai
  • Founder of Vora, 1,000+ travellers served
  • Hermes, a chief-of-staff agent on a Jetson Orin
  • Telecommunications engineering, UNI, Lima

01 · Capture

Thirteen domains, sized by use

Thirteen skill domains, sized by how often the work uses them.

The skill Cluster — curated domains of competence, not individual technologies. Size is derived from real incoming edges: how often a domain shows up across projects and roles, never authored by hand.

13 skill nodes, dot size = derived usage

  1. Cloud & deployment

    used 10×

    Shipping to managed runtimes — Cloud Run, Vercel and the AWS pieces that hold state.

    Google Cloud Run · Vertex AI · AWS · EC2 · S3 · RDS · Lambda · IAM · Vercel · Cloudflare · Firebase · Modal

  2. Frontend engineering

    used 8×

    Product interfaces in React and Next.js — state, streaming UI and dashboards people actually use.

    React · Next.js · TypeScript · Tailwind CSS · shadcn/ui · Zustand · Chart.js

  3. Multi-agent systems

    used 8×

    Designing agent graphs that plan, branch and recover — fan-out, barriers and human-in-the-loop.

    LangGraph · LangChain · LangSmith · ReWOO · HITL · Vercel AI SDK · Mixture-of-Agents

  4. Data & vector search

    used 7×

    Getting data in and finding it again — scrapers and connectors feeding Postgres and vector indexes.

    PostgreSQL · Supabase · pgvector · pgmq · MongoDB · DynamoDB · SQL · Apify · SerpAPI · Scrapling · CocoIndex

  5. Backend & APIs

    used 6×

    Async services that stream — FastAPI and Spring Boot behind agent pipelines and CRMs.

    FastAPI · NestJS · Spring Boot · SSE · asyncio · Pydantic · Java · Python

  6. LLM & RAG

    used 6×

    Grounding language models in real data with retrieval, embeddings and structured output.

    RAG · embeddings · pgvector · OpenRouter · Gemini · GPT-4o · Claude · Llama · Hugging Face

  7. 3D & WebGL

    used 2×

    Real-time graphics on the web — graph visualisation, scroll-driven cameras and shader work.

    Three.js · react-three-fiber · cosmos.gl · Spline · GLSL

  8. DevOps & containers

    used 2×

    Containers, pipelines and self-hosting — including hardening a box that runs unattended.

    Docker · Dokploy · GitHub Actions · Linux · Git · CI/CD

  9. Leadership & community

    used 2×

    Building the room — running hackathons, chapters and a podcast, and leading mixed teams.

    Scrum · Design Thinking · public speaking · event organisation · mentoring

  10. Edge & hardware

    used 1×

    Running inference off the cloud — agents and tiny models on Jetson, Raspberry Pi and Arduino.

    NVIDIA Jetson Orin · Raspberry Pi · Arduino · TinyML

  11. ML & computer vision

    used 1×

    Training and shipping vision models — anomaly detection, OCR and real-time counting.

    PyTorch · TensorFlow · OpenCV · MediaPipe · OCR

  12. Web3 & tokenization

    used 1×

    On-chain product work — asset tokenization and stablecoin rails with zero-fee on/off-ramps.

    Hedera Hashgraph · Base · viem · onchainkit · USDC · Solidity · Coinbase CDP

  13. Signal processing

    used 0×

    The telecommunications half — transforms, spectra and acoustic detection in the field.

    Matlab · FFT · IFFT · C/C++ · telecommunications

02 · Connect

Six roles that built the skills

Six roles, in order — where those domains were exercised.

The experience Cluster — where the domains above were actually exercised, in order.

6 experience nodes

  1. Vice Chair

    2025-01 → 2025-12

    Lead the Computational Intelligence Society at UNI, organizing workshops, courses, events, and projects.

    Led → Leadership & community

  2. Director of Leadership

    2025-03 → 2025-12

    Founded podcast Tech Talk Lead UNI interviewing Microsoft professionals and LATAM tech leaders.

    Led → Leadership & community · Built here: Tesorito — agentic cash-flow for Peru's micro and small businesses

  3. Frontend Developer

    2025-04 → 2025-10

    Frontend Developer responsible for building SaaS web applications to digitize and automate administrative and rapid-response processes in an emergency control and monitoring center. Led development of 4 interconnected systems that transformed manual operations into digital workflows.

  4. Full Stack Developer

    2025-10 → Present

    Full Stack Developer with experience designing, developing, and maintaining Web3 platforms focused on digital asset tokenization.

    Built here: ALTOKEN — real-estate tokenization on Hedera

  5. Software Engineer

    2026-05 → 2026-06

    Developed backend services with Spring Boot and PostgreSQL, contributing to the optimization of the company's CRM, lead-capture, and ERP systems.

  6. Entrepreneur in Residence (EIR) - AI Builder

    2026-06 → Present

    Design and deliver AI-native applications for enterprise customers while exploring the next generation of agentic AI systems, with mentorship from the Endeavor and Google ecosystems.

    Built here: Vera — a Chief of Staff agent for C-level leaders, Built here: Xertica-Education — AI content creation platform

03 · Build

What shipped

Eleven projects. Five are Case Studies; two open up into agent graphs.

The project Cluster. Five Case Studies get their own page; the rest ship as smaller cards.

11 project nodes

Case studies — Tier 1

  1. Case study

    Hermes — a self-hosted chief-of-staff agent, on the edge

    A chief-of-staff agent running 24/7 on a Jetson Orin Nano Super, orchestrating a Raspberry Pi and Claude Code without a cloud bill.

    NVIDIA Jetson Orin · Raspberry Pi · Claude Code · Docker · Dokploy

    Read the case

  2. Case study

    Sabueso — a multi-agent investigator for LATAM anti-corruption reporting

    An orchestrator fanning out to 6 parallel investigator agents, verified by a Mixture-of-Agents pass, to synthesise public-record research a journalist would otherwise do by hand.

    Next.js · Vercel AI SDK · cosmos.gl · FastAPI · LangGraph

    Read the case

  3. Case study

    Tesorito — agentic cash-flow for Peru's micro and small businesses

    A simplified, agent-assisted cash-flow and ERP tool for MYPEs, built for the Ideatech Hackathon and awarded 1st place with the LEAD UNI team.

    TypeScript · Turborepo · Supabase

    Read the case

  4. Case study

    Vera — a Chief of Staff agent for C-level leaders

    A reporting and second-brain agent that runs inside each leader's own Google Workspace: Forms capture the team's daily and weekly reports, Gemini summarises and consolidates them, and a Drive wiki builds itself from every report with hybrid embedding + lexical search. A Cloudflare Worker exposes it to Claude and ChatGPT as a 13-tool MCP server. Apps Script library, 421 tests.

    Google Apps Script · Gemini · Google Sheets · MCP · Cloudflare Workers

    Read the case

  5. Case study

    Vora — an AI travel agent that plans in under 60 seconds

    A multi-agent LangGraph system that turns a single message into a full itinerary — 8+ external APIs, a 1,000-video TikTok enrichment pipeline, and 1,000+ travellers served.

    FastAPI · LangGraph · LangSmith · Next.js · Supabase

    Read the case

Inside Sabueso — a multi-agent investigator for LATAM anti-corruption reporting

12 nested nodes, part-of Sabueso — a multi-agent investigator for LATAM anti-corruption reporting. Scrolling on descends through them and returns to the build Cluster on its own.

  1. 01

    load context

    Loads the target entity and the last claims from previous investigations so the planner does not rediscover them.

  2. 02

    plan

    Sabueso (Claude Sonnet 4.6, 4 prompt-cache breakpoints) turns the case into plan steps; a conditional fan_out edge sends one to each investigator in parallel.

  3. 03

    buscador

    El Buscador (ReWOO) resolves DNI and RUC identifiers and official-visit records via find_dni_record, find_ruc_record and search_manolo.

  4. 04

    tasadora

    La Tasadora (ReAct) reads JNE sworn-asset PDFs and checks them against SUNARP property records for patrimony discrepancies.

  5. 05

    contador

    El Contador (ReWOO) searches SEACE public-procurement releases per RUC and flattens each award into a claim.

  6. 06

    letrado

    El Letrado (ReWOO) queries the Legalize corpus for laws voted, sentences and vote-vs-interest conflicts.

  7. 07

    detective

    El Detective (ReAct) walks family, partner and board relations via find_relatives and SUNARP directories, emitting graph edges rather than claims.

  8. 08

    periodista

    El Periodista (ReAct) searches the press archive, snapshots pages in the Wayback Machine and follows up on the Twitter archive.

  9. 09

    collect

    Barrier after the fan-out: LangGraph converges here once every Send finishes and the reducer has concatenated all claims, edges and events.

  10. 10

    jueza

    La Jueza verifies every claim with a Mixture-of-Agents: 3 proposers in parallel against the primary source, then a Sonnet 4.6 aggregator calibrates the final verdict.

  11. 11

    synthesize

    Sabueso as editor-in-chief (Claude Opus 4.7, one call) writes the Markdown dossier: summary, findings, evidence, contradictions, next steps.

  12. 12

    persist

    Writes the investigation, its claims and edges to Postgres in one transaction — all or nothing.

Back out to Sabueso — a multi-agent investigator for LATAM anti-corruption reporting. See the full agent graph in the case study

Inside Vora — an AI travel agent that plans in under 60 seconds

15 nested nodes, part-of Vora — an AI travel agent that plans in under 60 seconds. Scrolling on descends through them and returns to the build Cluster on its own.

  1. 01

    classify intent

    Classifies the message intent with an LLM; a deterministic guard handles a pending HITL confirmation.

  2. 02

    extract preferences

    Extracts destination, dates, budget and travel style as structured output; merges with prior state.

  3. 03

    generate response

    Generates the conversational reply — exploration, one clarifying question at a time, or a confirmation prompt.

  4. 04

    confirm plan

    Builds the structured plan summary for the HITL confirmation widget, currency resolved by country.

  5. 05

    handle confirmation

    Processes the traveller's confirmation and unblocks the search pipeline.

  6. 07

    place searcher

    Searches points of interest against Google Places, a pool of roughly 100 candidates per trip.

  7. 08

    mobility searcher

    Searches flights and transit/drive routes in parallel, with per-airline deep links.

  8. 09

    accommodation searcher

    Searches Airbnb listings, check-in/out computed from the trip dates.

  9. 10

    build itinerary

    Builds the day-by-day itinerary with the LLM in adaptive batches, enriched with photos and coordinates.

  10. 11

    viator enricher

    Attaches bookable Viator tours to each day, ranked by proximity and category relevance.

  11. 12

    restaurant searcher

    Finds lunch/dinner restaurants per day, ranked by distance, rating and popularity.

  12. 13

    video enricher

    Attaches TikTok-derived videos via semantic search over a pgvector embedding index.

  13. 14

    extract refinement delta

    Extracts a structured delta from a follow-up change and determines its scope.

  14. 15

    handle refinement

    Applies the delta to state, clearing only what the scope requires so search re-runs selectively.

Back out to Vora — an AI travel agent that plans in under 60 seconds. See the full agent graph in the case study

More projects — Tier 2

  1. AGROX — crop anomaly detection

    A computer-vision platform for detecting crop anomalies in the field, awarded 2nd place at the LUPIC Hackathon.

    Python · OpenCV · PyTorch · FastAPI · React · Docker

    agrox-e963d.web.app
  2. ALTOKEN — real-estate tokenization on Hedera

    A regulated platform for fractional real-estate investment in LATAM: KYC/AML, fiat and crypto on-ramps, and holdings issued as Hedera Token Service tokens, read back through Mirror Node. Next.js 16 monorepo with an Express worker, Supabase Postgres as system of record and Payload CMS on MongoDB.

    Hedera · HTS · USDC · Next.js · Express · Supabase · MongoDB · Payload CMS · react-three-fiber

    www.altoken.io/en
  3. DepaseoX — a proptech advisor

    A proptech advisor built for the IA Hackathon Growt at UNI: property analytics and guidance behind a bilingual dashboard.

    React · Vite · TypeScript · Tailwind CSS · Supabase · Chart.js · i18n

    depaseox-ia-growt.vercel.app
  4. NUNA — a WhatsApp travel guide agent

    A travel-tech AI agent that guides a trip over WhatsApp, drawing on local knowledge. Joseph served as CTO; 1st place at UNI's Impact Start Competition, with an international bootcamp scholarship.

    WhatsApp · AI agents · LLM

  5. WINAYA — AI agents for Peruvian women entrepreneurs

    An AI agent platform helping Peruvian women entrepreneurs run their businesses, built for the Llama Impact Pan-LATAM Hackathon. Top-4 LATAM finalist.

    React · DynamoDB · Llama · FastAPI · LangChain

    winaya-lablab.vercel.app
  6. Xertica-Education — AI content creation platform

    An AI-powered educational content creation platform: FastAPI and Next.js over Supabase, generating text, images and video for course material.

    FastAPI · Next.js · Supabase · OpenRouter · gpt-image · Vertex AI · Veo · YouTube API · Google Drive API

    xertica-education.vercel.app

04 · Ship

What it won

Six recognitions, each read back off the project that won it.

The recognition Cluster. Each award's "won for" line is read back off the project that declared the edge.

6 recognition nodes

  1. Ideatech Hackathon (ESAN, OIM & KOICA)

    2026-07

    1st place with LEAD UNI for a technology proposal to strengthen the formalization and sustainability of micro and small enterprises (MYPEs). Prize: S/ 7,500. (07/2026)

    Won for: Tesorito — agentic cash-flow for Peru's micro and small businesses

  2. Halketon 2026

    2026-06

    Selected for Halketon 2026, an international social-impact hackathon in Buenos Aires — reported by UNI and the Peruvian government. Not listed on either CV YAML; recorded directly in the CV Map.

    No linked project

  3. NUNA

    2025-03

    Travel Tech startup incubated at ISC. (03/2025) Link

    Won for: NUNA — a WhatsApp travel guide agent

  4. Llama Impact Pan-LATAM Hackathon

    2024-12

    Top-4 finalist across LATAM in a hackathon organized by Lablab.ai and Meta. (12/2024) Link

    Won for: WINAYA — AI agents for Peruvian women entrepreneurs

  5. Empoderatech Hackathon

    2024-12

    Winner of a hackathon organized by the Ministry of Women and Vulnerable Populations (MIMP). (12/2024)

    No linked project

  6. LUPIC Hackathon

    2024-08

    2nd place in a hackathon organized by Seoul National University of Science and Technology and UNI FIEE. (08/2024) Link

    Won for: AGROX — crop anomaly detection

05 · Learn

Writing & talks

Press, a podcast and public writeups of the work above.

The writing Cluster — press coverage, a podcast, and public writeups of the work above.

4 writing & talks items

  1. UNI students selected for Halketon 2026

    2026-06

    Government and university coverage of the Halketon 2026 selection, an international social-impact hackathon in Buenos Aires.

    Press · gob.pe / UNI

  2. NUNA: the startup reimagining tourism with AI

    2025-03

    University coverage of NUNA, the WhatsApp travel-guide agent Joseph served as CTO on.

    Press · portal.uni.edu.pe

  3. Founded the podcast Tech Talk Lead UNI

    2025-03

    A podcast on Microsoft and LATAM tech leaders, founded while Director of Leadership at LEAD UNI.

    Podcast · LEAD UNI

  4. WINAYA project listing

    2024-12

    Public writeup of WINAYA on lablab.ai following the Llama Impact Pan-LATAM Hackathon.

    Press · lablab.ai

Vora travellers
1,000+
SJL residents served
1M+
Vora agent nodes
15
Hackathon wins
6

Get in touch

Reach me however you prefer

Email, LinkedIn, GitHub, or book a call directly — no form, no gatekeeping.

Booking link is a placeholder — swap it for a real Cal.com URL before launch.