Online Survey Portal with AI-Assisted Data Management

Turn a twice-yearly physician prescription panel from paper into a self-service digital process, and speed up drug reference consolidation after each data collection wave while keeping the quality of the dataset behind the reporting provided to pharmaceutical companies.

Specific goals:

1. Let physicians record appointments and prescriptions online at the point of care, including new drugs and brands missing from the reference.

2. Keep one golden drug reference while thousands of contributors keep extending it.

3. Replace manual post-production drug reference consolidation with an AI-assisted master data workflow run by one data manager, and substantially reduce consolidation time.

4. Make every change to production data reversible, gated and fully journaled.

5. Give medical experts a structured quality review of every submitted diary before data leaves the platform.

• Live end-to-end platform: paper replaced by one chain of digital entry, expert review, consolidation and publication.

• Scale per wave: 1,700+ physicians, about 80,000 patient visits, about 190,000 prescriptions.

• A multi-level drug reference graph with tens of thousands of entries kept as one golden reference while physicians extend it.

• Millions of snapshot rows processed in the sandbox over dozens of consolidation runs; every production change journaled and reversible.

• 100% of published merges pass a blocking integrity gate: no dangling references in production.

• Every physician entry kept in raw form: no data lost or corrupted while the model evolved.

• 250,000+ expert verdict transitions across multiple review levels.

• Drug reference consolidation time after each wave substantially reduced; run by one data manager.

• Web platform: React + TypeScript client, Node.js API, PostgreSQL.

• Data architecture: a raw data layer that is never overwritten, and a master data layer on top with sandbox snapshots, exact, fuzzy and structural matching, crosswalk mapping, an operations ledger for rollback and an append-only event journal.

• AI: large language model API for synonym expansion, grounded selection among existing golden records and clustering of new values; response cache and usage log.

• Quality control: multi-level expert review workspace with cascading verdicts and change history.

• Delivery: post-processing and handover to dashboards for partners and clients.

• Operations: hosting in Kazakhstan, nightly backups, scheduled consistency checks, monitoring dashboard and external uptime monitoring.

• Engineering: AI coding agents for design, code, review and documentation.

• Built on a live system: consolidation, quality review and export were designed and shipped while 1,700+ physicians entered data, through 100+ schema changes in production.

• Layered architecture set at the start: a raw layer that is never overwritten, and a consolidation layer with kept merged records, soft deletes and append-only journals enforced by the database.

• Graph-shaped reference: every merge is checked for cycles, self-links, leftover links and broken drug chains before commit.

• Safe change at scale: matching runs asynchronously beyond request time limits, mappings rebuild idempotently, concurrent data stewards work under dictionary leases.

• Grounded AI: a two-stage design so AI cannot invent reference records.

Initial situation. The client runs a physician prescription audit twice a year: more than a thousand physicians record their patients and prescriptions, and the aggregated data is provided to pharmaceutical companies. Until 2025 the questionnaires were on paper; operators retyped them into a local program and then into a database. After each wave the drug data was consolidated by hand: new brands, misspellings and incomplete drug chains were matched to the reference manually, which took about one month of post-production.

Implemented solution. An end-to-end chain from paper to published data: digital entry, quality review, consolidation, publication. Physicians enter diaries online and pick drugs through a guided step-by-step selection, adding missing values themselves. A submitted diary is locked and assigned to a medical expert automatically. Experts review each diary at several levels with cascading verdicts and corrections. The master data module takes a sandbox snapshot, proposes golden-record candidates by exact, fuzzy and structural matching, adds AI suggestions where spellings share no letters ("No-Spa" vs "Drotaverine"), and lets the data manager decide per value: replace, promote to golden, edit, reject or retire. Execution shows a human-readable impact preview, applies all decisions in one transaction and can be reversed.

Implementation specifics. Everything was built while the wave was live. The key decision, taken at the start, was a layered architecture: a raw data layer that keeps every physician entry as entered, and a consolidation layer on top that never overwrites it. The drug reference is a graph, not a list: a merge must not leave a dead end in the physician's selection path. We built blocking gates against broken links, a connectivity check against half-built drug chains, and a resolver that maps stale identifiers when a physician saves a prescription. AI helped on three levels: analysing the legacy structures and designing the target architecture, writing and reviewing the code (over 85% of commits co-authored by AI coding agents), and working inside the product as the data manager's assistant.

Changes after implementation. Physicians enter data at the point of care instead of on paper. Experts review structured diaries. One data manager consolidates a wave's new drug values, and consolidation time has been substantially reduced. Every change is journaled and reversible, and the client receives a ready export for its reporting.

Kazakhstan: a nationwide physician panel across major cities and medical specialties. The platform is hosted in Kazakhstan. Output feeds the client's analytics used by pharmaceutical companies in the Kazakhstan market. The architecture and master data workflow are country-agnostic: the same platform can be deployed for similar panels in other markets. This means reusing the solution, not transferring data or extrapolating results to other countries.