luciel.dev

Jhamil Peña

AI / ML engineer · Lima

I turn manual processes into systems that run.

Open the labGet in touch

Tour
Sacred Valley
Date
14 Nov 2026
Travellers
4 adults
Pickup
07:00, hotel
Total
S/ 1,240.00
Deposit
50% paid
Sold by
Front desk
Status
Confirmed

Built to be rebuilt from git.

output
static HTML, rendered by Astro at build time
client js
< 2 kB per page: the dimension switch and the lab replay
fonts
self-hosted, nothing loaded from another domain
runtime
nginx in its own container, behind Traefik
deploy
rebuilt from the repository on every release

Two systems, built for real use.

tours

Live · login only
Context

A tour agency and hotel in Cusco ran sales, commissions and settlements on Excel and VBA macros. A deleted row left no trace.

Outcome

A real client uses it every day. It has no public demo, so the source code is the showcase.

Approach

A double-entry ledger from day one: every sale posts a balanced entry, and Python checks the balance inside the same transaction.

System

Next.js front end, FastAPI API, SQLite in WAL mode, role-based access and an audit log with before and after values.

Accreditation assistant

In development · faculty project
Context

The faculty's accreditation process needs large volumes of unstructured documents analysed and audited.

Outcome

On repeated queries, prompt caching cut the tokens sent for the system prompt by about 90%. Built with a classmate under faculty supervision.

Approach

An LLM assistant on Amazon Bedrock with multi-phase batch sessions, live monitoring and automatic recovery after failures.

System

FastAPI, async PostgreSQL, role-based auth, persistent rate limiting, S3 backups and AES-256-GCM encryption.

Hundreds of documents. Audited in minutes.

The accreditation assistant, opened up so you can try it: pick a sample document and get an audit where every finding points to the page it came from.

Open the labToday it runs on synthetic samples. Next: your own file, processed on your device.

syllabus-data-structures.pdfCourse syllabus · Data Structures · 6 pages
  1. Learning outcomes are statedMetp. 1
  2. Outcomes map to program objectivesMetp. 2
  3. Assessment methods are describedMetp. 3
  4. Student work is evidencedMissing—
  5. Improvement actions are recordedPartialp. 6
  6. Instructor qualifications are documentedPartialp. 1

  • Ask about my work

    A small model, running on your device, that knows what is on this site.

    Coming soon

  • Chess against a model

    A board, an opponent, and the reasoning behind each move.

    Coming soon

read → convert → analyze → audit → report

  1. 01readthe PDF is split into pages
  2. 02convertscans become text, tables are kept
  3. 03analyzean LLM works through it in phases
  4. 04auditeach criterion is checked against the text
  5. 05reportevery finding points to its page

Planned: the model runs in your browser on WebGPU, with a WASM fallback. Sample documents first, your own file as an option.

Who builds this.

I'm a Computer Science student at Universidad Peruana de Ciencias Aplicadas (UPC) in Lima, Peru, focused on data science, machine learning and generative AI. I work with Python (pandas, scikit-learn, PyTorch, TensorFlow), LangChain, SQL and Docker.

Right now I'm building, with a classmate and under faculty supervision, an LLM-assisted system on Amazon Bedrock that analyses large volumes of unstructured documents for the faculty's accreditation process. It uses a FastAPI backend, async PostgreSQL, multi-phase batch processing and prompt caching.

Before that I interned in Processes & Technology for Wealth Management at Credicorp Capital (Power BI dashboards, SQL optimization, Databricks/Spark migration support, SSIS ETL). I have also led student tech communities as President of IEEE CIS UPC and Project Director at IEEE ComSoc UPC.

From a café in Cusco to production systems.

  1. 2017 – 2022

    Basílica Café, Cusco

    Freelance tech solutions: automated demand reports for management and an ETL from the sales system into a central database.

  2. 2022 – now

    Computer Science, UPC

    Data science, machine learning and generative AI.

  3. Oct 2024 – Jun 2025

    President, IEEE CIS UPC

    Led the chapter's project portfolio and its industry partnerships.

  4. Jul – Sep 2025

    Project Director, IEEE ComSoc UPC

    Led a multidisciplinary team building an IoT inventory system with a microservice architecture.

  5. Oct 2025 – Mar 2026

    Intern, Credicorp Capital

    Processes and technology for Wealth Management: Power BI dashboards, SQL optimisation, Databricks migration support and SSIS ETL.

  6. Mar 2026 – now

    Full-stack & GenAI developer, UPC

    The accreditation assistant above.

Measured, not projected.

Each figure names where it comes from.

1client running in productiontours.luciel.dev
144automated tests on that app68 vitest · 76 pytest
~90%fewer system-prompt tokens on repeated queries, with prompt cachingaccreditation system · CV
< 2 kBclient JavaScript on any page of this sitethe dimension switch and the lab replay

What I build with.

ML & deep learning
Pythonpandasscikit-learnPyTorchTensorFlow
LLMs
Amazon BedrockLangChainprompt caching
Data
SQLPostgreSQLDatabricksPower BI
Systems
FastAPIDockerTraefikAWS

How this site is built.

  • Real tools, not demos

    Each project is a tool that solves a real problem, not a demo with nothing behind it.

  • One subdomain per tool

    Every tool lives on its own subdomain, with genuine content written around it.

  • Reproducible infrastructure

    Everything runs on infrastructure versioned in git: Traefik and Docker Compose, one container per app.

Every tool on the platform.

  • toursLive
  • rtkComing soon
  • graphComing soon
  • hackathonsComing soon

See details on the Projects page

Got a process that eats hours?

Tell me how it works today. If AI can take it off your hands, I'll show you how.

Get in touch

Read the source.

This site and every tool it lists live in one public monorepo.

github.com/Jaed69/luciel-platform ↗