Martín Ramírez Espinosa

Selected projects

Work that can be inspected, reproduced, and questioned.

Six projects selected for the engineering decisions they expose—not for the number of technologies they contain. Status and limitations remain visible alongside the result.

RAGdoll research with receipts project banner

Research systems · AI

RAGdoll

Scholarly research, with receipts.

A terminal workspace that turns ambiguous research questions into curated, cited literature dossiers while preserving human approval, retrieval provenance, evidence level, and exact supporting passages.

  • Separate approval boundaries for search, paper curation, and open-full-text acquisition.
  • A recorded run indexed 307 page-aware passages and produced a seven-section dossier.
  • 98 offline tests, strict typing, security-focused adapters, and a deterministic no-key demo.
  • Python
  • Textual
  • SQLite FTS5
  • OpenAI
  • Ollama
OpenAI Build Week submission · v2.2 release candidate
Read case studyRepositoryDocumentationVideo
Sealed-test bits per character results for smaLLM tokenizers

Language models · Reproducible research

smaLLM

A GPT-style language-model lab built from scratch.

An inspectable PyTorch implementation spanning tokenization, causal attention, training, generation, controlled baselines, sealed evaluation, and reproducible experiment artifacts.

  • A preregistered panel matched the character control within 0.40% of ByteBPE512's parameter count.
  • ByteBPE512 won eight of nine matched sealed-test comparisons; the Douglass reversal remains documented.
  • Completed at version 1.0 with strict typing, 182 tests, 90.49% branch coverage, and a five-minute CPU demo.
  • Python
  • PyTorch
  • Transformers
  • Experiment design
Completed research artifact
Read case studyRepository
insIGht local Instagram export import interface

Desktop product · Privacy

insIGht

Instagram relationship analytics without an Instagram login.

A local-first desktop application that turns official Instagram exports into searchable relationship snapshots without scraping, telemetry, cloud processing, or unofficial APIs.

  • Rust owns archive parsing, validation, SQLite persistence, reports, and native filesystem access.
  • Defensive import limits bound paths, file counts, individual files, and aggregate decompression.
  • Typed Tauri commands connect a React interface to immutable local snapshots.
  • Rust
  • Tauri
  • React
  • TypeScript
  • SQLite
Cross-platform desktop application
Repository
Top render of the AudioLab TLV320AIC3104 codec daughterboard

Electronics · Embedded DSP

AudioLab

A mixed hardware, firmware, and host-software audio platform.

A NUCLEO-F429ZI and TLV320AIC3104 platform for USB audio, codec control, real-time DSP, measurement, and validation across explicit hardware and software boundaries.

  • Implemented Altium daughterboard with project-owned libraries, schematics, PCB outputs, renders, and a 3D model.
  • Architecture separates STM32 firmware, hardware design, host tooling, specifications, and verification artifacts.
  • The repository labels proof-of-concept firmware honestly while the final codec contract remains in progress.
  • C
  • C++
  • STM32
  • USB Audio
  • Altium
Hardware implemented · firmware boundary evolving
Repository
Editorial cover representing SIAlytics academic analytics

Data product · Browser automation

SIAlytics

Academic progress students can inspect for themselves.

A defensive local workflow that reads an authorized UNAL academic-history session and produces an Excel workbook for semester trends, curriculum progress, PAPA, and PAPPI.

  • Sign-in remains manual; sensitive session material and authenticated pages are never retained.
  • Unknown structures, incomplete pagination, expired sessions, and ambiguous empty states fail closed.
  • The end-to-end extraction was validated against a real SIA session on June 28, 2026.
  • Python
  • Playwright
  • OpenPyXL
  • Jupyter
Validated student tool · Spanish interface
Repository
Editorial cover representing kernels, learning theory, and model assessment

Mathematics · Coursework archive

Machine Learning Theory

The mathematical foundation behind the systems work.

A navigable UNAL course archive covering statistical learning, regression, kernels and RKHS, manual neural networks, model assessment, optimization, and time-series experiments.

  • Topic-oriented navigation preserves the original course structure without presenting coursework as a product.
  • Selected notebooks cover RKHS mean embeddings, manual backpropagation, diagnostics, and experimental comparison.
  • The repository is explicitly archived as a readable academic record.
  • Mathematics
  • Python
  • Jupyter
  • LaTeX
Archived academic record
Repository