The problem
CRM and billing exports rarely agree on names, formats, or identifiers. Before someone can answer a business question, they often need to understand the schema, match records, and inspect the exceptions. Meldify brings those tasks into one workflow.
THE RECONCILIATION WORKFLOW
My role as founder
I founded Meldify and lead its product development, from the reconciliation workflow and AI assistant to the interface people use to review results. My work connects the product decisions with the implementation across the backend and front end.
What I built
The reconciliation pipeline maps source columns into a canonical model, proposes candidate record pairs, and uses Fellegi–Sunter scoring with expectation-maximisation to classify matches. One-to-one assignment resolves competing matches. Saved mappings and recipes make repeated imports easier to reproduce.
Alongside that pipeline, a conversational assistant can query datasets, explain computed results, propose changes, and prepare chart specifications. FastAPI serves the backend; a React and TypeScript interface makes the results and review steps accessible.
Where AI helps — and where code does the work
Schema mapping combines deterministic candidates, TypeSafe/Jev judgments, and an LLM fallback. The assistant uses a bounded tool-calling loop and schema-constrained tool arguments. Polars computes results from the data rather than asking a language model to estimate them from a sample.
Chart generation follows the same pattern: the model proposes a chart specification, then the application computes the values. Proposed edits are previewed; the assistant does not get a direct apply-changes tool.
THE ASSISTANT WORKFLOW
Engineering decisions
- Human-reviewed mappings: the person can inspect proposed schema mappings before reconciliation.
- Evidence-backed answers: the assistant refers to computed results, with evaluation checks against ground truth.
- Revision-aware proposals: proposed edits are tied to a dataset revision so they can be checked before application.
- Explicit boundaries: tool budgets, dataset authorization, and restricted operations limit what the assistant can attempt.
What the project demonstrates
Meldify connects AI orchestration to a substantial data workflow: mapping, entity resolution, analysis, charting, and review. It shows how I approach the interface between model output and deterministic software.
Meldify is the product I founded and am actively developing. The case study describes implemented workflows, not customer adoption or guaranteed business outcomes. Schema assistance can use column headers and sample values; matching and calculations run in application code.