Home / Projects / Construction Data Onboarding Workbench
Construction Data Onboarding Workbench
Construction Data Onboarding Workbench is a project by Akash Renuka Ashok (AI Engineer (independent prototype)). An internal delivery tool for construction data migration: messy customer CSV in, canonical Company, Worker and Certification entities out, with a mandatory human approval gate before anything is exported.
- Role
- AI Engineer (independent prototype)
- Category
- Forward-Deployed Prototype
- Stack
- React, TypeScript, FastAPI, SQLite, Ollama (llama3.1), OpenAI/Anthropic/Gemini adapters, pdfplumber
The problem
Customer onboarding data arrives inconsistent, duplicated and incomplete, and a migration that silently merges or guesses records creates errors nobody can trace back to the source.
What Akash owned
Designed and built the full prototype solo on synthetic data: validation, deduplication, review interface, exports and the LLM drafting layer. Not affiliated with any commercial product.
Highlights
- Deterministic-first by design: validation, ABN checksums, date parsing, deduplication and export eligibility are ordinary testable code, and the LLM only drafts customer-facing summaries and explains duplicate evidence
- Expanded a deliberately messy 28-row customer file into 62 validated entities, surfacing 26 blocking errors and 10 warnings, each traceable from exported field back to source row
- Deduplicated with deterministic normalisation and exact matching first, surfacing fuzzy matches as candidates for human approval rather than merging them
- Exports import-ready and excluded files, an exception report, an audit log and an implementation summary
- Added a secondary flow extracting structured fields from machine-readable Safe Work Method Statement PDFs, with source evidence per value
- Runs fully local on Ollama by default, with adapters for hosted providers
Outcome
Demonstrated a deterministic-first approach to AI-assisted data onboarding, where every exported field traces to its source row and the model never merges, approves or produces a number.
More work by Akash Renuka Ashok:
Last updated