Health-tech product / Engineering case study
A mobile app, admin panel
and calculation services for physicians
A physician-focused platform combining reference-guided AI and TUS / DHY preference tools. I took end-to-end technical responsibility across the mobile app, admin panel, backend and background computation systems.
01 / The product
BTM helps physicians consult forensic-medicine references and compare positions during TUS and DHY placement rounds. I built both experiences into the mobile app, with separate data models and access rules for each.
Alongside the mobile app, I built APIs, database models and admin tools for document reviews, official data imports and calculations. The mobile app and admin panel use the same backend, with different permissions for users and administrators.
I built the mobile chat and backend flow that bring case findings, conversation history and a reference document into a controlled LLM request. Access, conversation ownership and usage limits are checked on the server.
Gemini · Edge Functions · Mobile chatI built position discovery, filters, map views, historical results and term-specific preference lists. TUS positions are modeled as hospital–specialty pairs, while DHY keeps its distinct lottery and preference rules.
TUS / DHY · History · Preference managementI developed the Next.js panel for document review, position management, moderation and computation tracking, connecting operator decisions to permission-checked backend actions and the mobile experience.
Next.js · Server permissions · Operations02 / BTM AI
I connected Gemini to the reference document and conversation history, with instructions to stay within the source, acknowledge missing information and ask for clarification when findings are ambiguous. The application controls document preparation, authorization and persistence.
I also developed an SSE response path and mobile chunk handling for incremental answers, including partial-response and interruption handling. Reference versions and provider-file validity are checked before reuse.
For ÖSYM result documents, text PDFs are parsed directly; scanned content can use Gemini for transcription. A deterministic parser then interprets and validates the score fields. This keeps the data rules inspectable even when the source needs OCR.
BTM AI provides information from a reference guide to support a physician’s assessment. I also built streaming responses, which can be enabled through the app’s release settings.
03 / Prekura computation
I built separate engines for TUS score-based placement and DHY lottery scenarios. Heavy computation runs in background services using PostgreSQL job queues and Node.js worker threads.
I save the inputs for each run and use a fixed random seed so the same calculation can be repeated. Input and output hashes help identify the data used. TUS scores use an integer representation to avoid rounding differences. These checks make calculations reproducible; they do not establish how accurately a model predicts real placements.
Admin-created runs freeze the candidate, preference and position inputs. Node.js worker threads process TUS placements or DHY scenarios, prepare results in stages and publish the completed run.
Preference and profile changes create new revisions. I added job ownership, heartbeats, retries and stale-job handling so an older calculation cannot replace the result of newer input. Final runs receive priority.
A separate worker follows DHY’s lottery model. Rapid input changes are grouped before computation, with leases and heartbeats coordinating work and allowing final calculations to take precedence.
Shadow, pilot and enabled modes separate computation from user publication. A notification outbox rechecks the active term, release mode and user permission before sending, with deduplication and invalid-token cleanup.
I built a separate system for testing different candidate distributions. Each run records its model version and results, and cannot start if the required model files are missing or invalid.
Research and testingI designed the probability calculations to run independently of whether their results are shown to users. The national model using synthetic candidates remains separate for research and testing.
04 / The admin panel
I built the Next.js admin around the work that happens after a user submits a document, reports a message or changes a preference. Review queues, filters, decisions and calculation details turn those events into manageable operational workflows.
The panel lets reviewers filter pending documents, open them through temporary links and record their decisions. The server checks their permissions before granting access.
Administrators can review reported comments and messages and manage blocked users. Blocking rules are also enforced by the backend.
The panel lets administrators manage positions for each term and inspect calculation inputs, progress and results. Long-running calculations continue in background services.
For important changes, I record who made the change and the previous and updated values of selected fields. Administrators can use this history to investigate past decisions.
I enforced role and action permissions on the server, alongside user ownership and row-level data access controls. Document review receives temporary access only after authorization, keeping sensitive operational work behind explicit checks.
05 / Official data
I developed import and validation tooling to move official position data through source capture, parsing, staging, review and publication. Existing position identifiers and history are preserved so new data does not break preference links.
Source hashes, import-run records and validation reports make the publication traceable. The workflow combines scripts with human review before approved data reaches production.
Save the official source and the information needed to identify it.
Prepare the records in a staging area for review.
Review record matches, totals and existing preference links.
Publish the approved data and keep a record of the import.
I imported 3,825 position records covering 15,501 available places for the 130th DHY round.
06 / Architecture & delivery
Authorized data access + server operations
Application-controlled integrations + background computation
I implemented subscription entitlement handling for duplicate and out-of-order RevenueCat webhook events, protecting access and usage limits from delivery ordering issues. Push delivery uses deduplication records and device-token cleanup.
I worked across EAS builds, OTA configuration, release maintenance and Sentry integration. Parser, placement-engine and subscription-event tests support the underlying rules, while worker status and heartbeat records expose long-running activity.
07 / The stack
For BTM, I developed the mobile app, admin panel, AI integration and background calculation services. I also handled the data models, subscription integration and release maintenance.
From product flow to background job