Artifact Rescue

Restore abandoned spreadsheets into modern applications—and prove every recovered rule before migrating real data.

Comment

Inspiration

Small organizations often depend on a spreadsheet or desktop tool built by one volunteer or employee. When that person leaves, the formulas, exceptions, and institutional knowledge can disappear with them. Replacing the interface is easy. Proving that the replacement still makes the same decisions is the hard part.

Artifact Rescue is a restoration studio for that dying software.

What it does

A user uploads an .xlsx workbook. Artifact Rescue inspects its sheets, columns, formulas, comments, hidden content, supported behavior, and conflicts, then produces a Recovery Compatibility Report and an evidence-backed Rule Ledger. Every recovered rule cites its source cells, original formula, confidence, affected records, and confirmation status.

GPT-5.6 turns ambiguous workbook evidence into focused clarification questions and can propose constrained formula changes from a reviewer’s explanation. The model is intentionally not the trust boundary: unsupported behavior stays unknown, every proposed rule must pass schema and grammar validation, and a deterministic fallback keeps the workflow usable if the model is unavailable.

The recovered blueprint generates a modern application with real fields, role-gated data entry, searchable tables, calculations, reports, and CSV export.

The defining feature is the Proof Room. It runs sample, boundary, and judge-created scenarios through both the recovered legacy rules and the restored rules. Matches turn green, differences turn red, and untested behavior remains visibly unknown. A reviewer can explain a mismatch, approve the constrained change, and rerun the entire regression suite.

Before migration, Artifact Rescue maps real uploaded rows, finds missing values and duplicates, previews possible loss, imports transactionally into normalized Supabase tables, reconciles checksums, and supports rollback. The final restoration package contains the rebuilt application, migrated records, operating guide, machine-readable certificate, and PDF certificate.

How I built it

The product is a Next.js application deployed on Vercel, backed by Supabase Postgres. ExcelJS recovers workbook structure and formulas. A constrained formula interpreter executes arithmetic, comparisons, IF, Boolean logic, aggregates, conditional aggregates, rounding, and constrained VLOOKUP/XLOOKUP behavior without food-bank-specific code. The Vercel AI SDK connects to GPT-5.6 through OpenRouter for structured clarification and rule proposals. Zod validates model output. Playwright exercises the complete desktop and mobile restoration journey, and GitHub Actions runs linting, self-checks, production builds, and browser tests on every push.

Codex was my engineering partner across the build. It helped turn the restoration-studio concept into the workbook recovery pipeline, generic interpreter, Proof Room, migration architecture, regression suite, security hardening, deployment workflow, and Remotion demo. The key decisions remained explicit throughout: recovered behavior must cite evidence, uncertainty must not be invented away, and no migration becomes permanent before deterministic verification and human approval.

Challenges

The hardest problem was avoiding a convincing but untrustworthy demo. It would have been easy to hardcode one food-bank workflow or let an LLM calculate expected answers. Instead, I built one generic pipeline used by food-bank, animal-shelter, and clinic-scheduling workbooks, kept numerical proof deterministic, and made unsupported behavior visible.

Formula recovery also required a deliberate compatibility boundary. Artifact Rescue supports a useful subset and reports everything outside it instead of silently guessing.

Accomplishments

  • Evidence provenance from every rule back to workbook cells and formulas
  • A generic old-versus-new Proof Room with branch coverage
  • Judge-created scenarios and regression reruns
  • Three domains through the same recovery pipeline
  • Transactional migration, checksum reconciliation, and rollback
  • A restoration scorecard and certificate that distinguish verified from unknown behavior
  • A working public deployment judges can test immediately

What I learned

AI generation becomes much more useful when it is surrounded by evidence, deterministic checks, explicit uncertainty, and reversible actions. The important question is not “Can AI rebuild this?” It is “Can the organization prove what was preserved?”

What's next

Next I would add broader Excel function coverage, screen-recording-assisted workflow observation, multi-user review queues, richer generated dashboards, and connectors for Access databases and older desktop tools.

Generation is not the trust boundary. Proof is.

Built With

Share this project:

Updates