Web Platform for Automated Technical Report Production

This project is a web platform for an environmental engineering consultancy: it replaces a scatter of spreadsheets, a legacy desktop calculation tool and manual assembly in Word with a single integrated system that produces their technical reports end to end. The governing principle: every number is produced by a deterministic calculation engine. Language models only write prose and assemble tables. No value in the final report is ever generated by a model. Scope of work Platform, ingestion and data model Web dashboard where technicians create projects, upload inputs, follow what the system is producing, correct or approve it, and retrieve the final output Uploads go directly to object storage, versioned per project, report and version, following the client's internal coding scheme, in a closed circuit Parsers for the field template (13 sheets: sampling points, lithology, piezometers, water levels, campaigns, soil and water samples, purges) and for lab result files, with strict schema validation, value normalisation and loud failure on unexpected formats A reference-value database as a first-class extensible entity: 52 tables (42 soil, 10 groundwater) covering 230 substances, versioned so that a report always reproduces under the rules in force when it was written, and open to site-specific tables the client adds themselves Deterministic calculation engines (Python, fully unit-tested) Contamination — comparison of lab results against the selected reference table, sample classification, standard result tables with the client's colour coding, including only substances measurable above the quantification limit in at least one sample Hazard — threshold evaluation under the applicable waste regulation, automated for the known contaminant families and flagged for human completion outside them Admissibility — correlation between base and admissibility samples proposed by the system and approved by a technician before the run continues Volumes — soil volume by category, mapped to disposal routes and unit prices; fed as validated input first, with full computation from topography, mask and borehole data as a later milestone Quantitative risk — reconstruction of a Tier 2 risk model from public standards and the client's own parameters: direct contact, indoor air via Johnson & Ettinger, outdoor volatilisation and particulates via the ASTM surface models, editable receptor exposure factors, reverse calculation of site-specific target levels, and a generated appendix matching the detail of the legacy tool Report generation Standardized text combined with model-written variable sections, in the client's institutional terminology Programmatic assembly of complex tables and appendices, with dynamic annex numbering, output as DOCX/PDF Bibliography handling that separates fixed references from site-specific ones Visible flags in the document marking exactly what a technician must review or fill before delivery Controls and validation Guard rails on model output and monitoring of usage costs, so monthly spend stays predictable and visible Full audit trail separating what the system generated from what a human approved, with provenance from every result back to its source file Numerical parity against the legacy tool, demonstrated on real historical cases within an agreed tolerance. This is the system's acceptance test, not an afterthought. Explicitly out of scope Any calculation performed by a language model Final sign-off — a client technician approves every report before it leaves the platform; the system produces, it does not certify Submission to regulators, licensing workflow, invoicing, CRM, and any portal for the client's own customers Delivery Six milestones, two-week cycles, tracked in Linear. Each milestone has written scope and acceptance criteria agreed before work starts. The build is followed by training for the end users, a hypercare period after go-live, and ongoing monthly maintenance.