Claude Code Hackathon Top 3: How a Turkish Doctor Won Gold with AI‑Powered MedKit

MedKit – Voice‑Driven Clinical Simulation

Bedirhan Keskin, a medical doctor turned software engineer, built a fully voice‑driven medical simulation that runs in a browser. The system uses streaming speech‑to‑text and text‑to‑speech so the user speaks and the AI patient replies in real time.

Persona behavior – The AI patient hesitates, asks follow‑up questions, and shows anxiety, implemented via a prompt‑cached persona rather than a generic chatbot.

Complete diagnostic loop – The workflow gathers patient history, orders laboratory tests, reads imaging results, writes prescriptions, and generates discharge notes.

Three‑dimensional scoring rubric – Data Gathering, Clinical Management, and Interpersonal dimensions are rated from “excellent” to “clear fail”. Each feedback item cites an actual clinical guideline (NICE, ESC, AHA, GINA, GOLD); the author enforces “no fabricated citations”.

Two usage modes

ER mode: multiple parallel beds, time‑pressured triage, simulated lab turnaround times.

Polyclinic mode: single‑patient flow with a Three.js‑rendered 3D clinic, supporting adult and pediatric cases.

Online demo: https://medkit‑app.vercel.app

Wrench Board – Board‑Level Diagnostic Workbench for Microsoldering Technicians

Alexis Chapellier created a workbench that consolidates scattered repair knowledge and provides AI‑assisted diagnosis.

Knowledge Factory – Four Claude roles (Scout, Registry, Writers, Auditor) collaborate for about two minutes to produce a certified knowledge package. Three parallel Writer agents share a prompt cache to reduce cost.

Schematic Ingestion – Opus 4.7’s vision model reads PDF schematics page‑by‑page, performing native visual reasoning (not OCR) and compiles a queryable ElectricalGraph with inferred net‑list order.

Diagnostic Agent – Runs on Anthropic Managed Agents with four layers of memory: global patterns, reusable playbooks, device‑specific packages, and session notes. Model tier switches per workload: Opus 4.7 (deep), Sonnet 4.6 (normal), Haiku 4.5 (fast). Custom tool definitions reside in api/agent/manifest.py.

microsolder‑evolve nightly loop – Four pipelines (deterministic simulator, schematic compiler, vision channel, diagnostic agent) execute automatically each night. The loop runs benchmarks against an oracle; successful changes are committed, failures trigger rollback.

Anti‑hallucination design – The agent is prohibited from fabricating component reference designators; every emitted ref‑des must be retrieved from a tool query, and a server‑side sanitizer blocks unverified tokens before UI display.

Boardview parsers – From‑scratch parsers for twelve formats (.kicad_pcb, .brd, .brd2, .asc, .bdv, .bv, .cad, .cst, .f2b, .fz, .gr, .tvw) enable easy addition of new formats.

Code repository: https://github.com/Junkz3/wrench-board (Apache 2.0, demo uses MNT Reform board, CERN‑OHL‑S hardware)

Maieutic – Intent‑First Coding Assistant

Paula Vasquez‑Henriquez built a tool that requires students to describe in natural language what they intend to build and why before the AI unlocks the code editor.

The workflow forces a metacognitive step: the user articulates the problem and design rationale, the LLM confirms understanding, then the code editor becomes available. This counters “vibe coding”, where a single prompt yields large code snippets without learning.

Author Reflections

Domain expertise × Opus 4.7 agent capability = immediate‑value tool

All three top projects were created by practitioners (doctor, microsoldering technician, educator) rather than large‑tech engineers, demonstrating that deep domain knowledge combined with Opus 4.7 agents yields instantly useful applications.

Each project pushes agents to their limits: MedKit uses a prompt‑cached persona for realistic patient dialogue; Wrench Board integrates 36 custom tools, four‑layer memory, and tiered model selection; Maieutic treats the LLM as a thinking coach rather than a pure code generator.

Anti‑hallucination safeguards are essential in error‑intolerant domains. MedKit’s mandatory citation of real clinical guidelines and Wrench Board’s refusal to fabricate component IDs illustrate that Opus 4.7 agents can operate safely in medical and electronics repair contexts.

“Let’s see where this journey will take us, but I genuinely look forward to being one of the people shaping the future of medical education.” – Bedirhan Keskin