Crafter Research / Lima, Peru
We turn public data, legal and civic corpora, and model evaluations into reproducible artifacts, datasets, registries, and accountability tools.
We work on source-backed datasets, scorecards, registries, evaluation artifacts, and small tools tied to public-interest questions.
Messy government and civic sources become reusable, versioned, and inspectable datasets.
Fragmented public material becomes structured source trails, corpus cards, and research inputs.
Evaluation work should separate correctness, abstention, hallucination, and evidence quality.
Five current projects with their released surfaces, supporting evidence, verification date, and known limits.
Official earthquake, accelerometric, and volcanic data from Peru and Colombia, exposed with traceable sources.
Operational history begins in July 2026, so long-term reliability is still unproven.
An agent-first CLI for supervised Peruvian tax workflows, with explicit safety controls and audit trails.
Several workflows depend on government portals whose interfaces and availability can change without notice.
An open observatory for Latin American cultural benchmarks that separates correctness, abstention, and hallucination.
The judge calibration is in-sample, and the public benchmark datasets have no contamination analysis.
A versioned corpus of Peruvian law and the open engine that produces, audits, and serves it.
The engine is pre-1.0, and the current refresh cadence is not yet documented as an operating commitment.
A census-wide, reproducible index of the digital surface Peru's municipalities declare on gob.pe.
Two scored snapshots demonstrate one recurrence, but not yet a stable monthly operating history.
Question, sources, artifact, review, documentation, release.
A public-interest question tied to a real source, corpus, or model behavior.
A source trail with URLs, dates, limitations, and evidence quality notes.
A dataset, eval, registry, scorecard, CLI, or accountability tool.
Check reproducibility, source quality, and limitations before claims harden.
Shareable notes that do not require private vault access to understand the work.
Promote, hold, or reshape the project based on evidence rather than excitement.
Projects can move at different speeds, but the output should stay inspectable and reusable.
Public sources made easier to cite, refresh, verify, and reuse.
Source trails, hashes, refresh logsFragmented law, records, and civic text turned into reusable research inputs.
Corpus cards, coverage notesEvaluation artifacts that make model behavior and evidence quality easier to inspect.
Rubrics, runs, agreement checksFocused tools that help people inspect public systems without hiding the source trail.
Registries, CLIs, dashboardsWe sampled one municipal domain from each of 100 municipalities. The useful result was not a new ranking, but a stricter boundary around what one observation proves.
Muniscan mapped the public digital surface of Peru's municipalities. Its most useful result came when the second scan failed.
We built a hybrid-RAG assistant over 21,000 Peruvian legal norms. Then we measured it, and the conclusion moved three times as the evaluation got more rigorous.
Every serious claim should point to a dataset, eval, source card set, registry, run, or public method.
Public-facing work should be source-backed, limitation-aware, and reproducible enough to inspect.
Project context should stand on its own without requiring access to private planning notes.