Dae Hyeon Kim RESEARCH 한국어

NeuroSync

Combined Korean LLMs with evidence verification to turn intake conversations into structured clinician handoff records.

GitHub: neurosync
Role
Method design and core implementation
Contribution
Implemented multi-agent orchestration, retrieval and evidence verification. Integrated the patient app, clinician web interface and backend.
Period & affiliation
2026.03 ~ Present NeuroSync

01 / RESEARCH DETAILNeuroSync: intake, retrieval and clinician handoff

Challenge
Free dialogue must populate clinical slots without allowing unsupported values into the handoff record.
Approach
A rule-based state machine runs 10 specialist agents. Scales use deterministic scoring, and final records undergo source and evidence checks.
Contribution
Implemented orchestration, RAG and verification logic and integrated the patient app, clinician interface and backend.
Intake, retrieval, scales, longitudinal analysis and evidence verification
Intake, retrieval, scales, longitudinal analysis and evidence verification
View full-size figure
Prototype handoff report: current state, PHQ-9 and risk level
Prototype handoff report: current state, PHQ-9 and risk level
View full-size figure
Models, retrieval sources and service stack
Component Implementation scope
Korean models Solar Pro 3, K-EXAONE, A.X K1
Retrieval corpus 1,248 clinical cases, 1,789 QA entries
Services FastAPI, PostgreSQL, Expo, Next.js
Evaluation conditions and scopeAI Champion intake prototype for preparing clinician handoff records; it does not make diagnoses.

02 / RESEARCH DETAILNeuroSync: output reliability and inference cost

Verification results
Verification blocked 337 unsupported candidates before final output and improved slot coverage and scale assignment.
Remaining issue
The pre-filter unsupported-candidate ratio rose from 46.4% to 55.2%, exposing limitations at the generation stage despite final-output verification.
Next steps
Further work concerns reducing calls and tokens and obtaining specialist assessment and clinical proof-of-concept evidence.
Structured intake performance (%) Single backbone: Scale assignment 23.8, Slot coverage 55.6; NeuroSync: Scale assignment 73.0, Slot coverage 98.4 0 25 50 75 100 23.8 73 Scale assignment 55.6 98.4 Slot coverage Single backbone NeuroSync
Structured intake performance (%)
Unsupported final values (count) Final output: Single backbone 39.0, NeuroSync 0.0 0 12 24 36 48 39 Single backbone 0 NeuroSync Final output
Unsupported final values (count)
Cumulative waiting time per turn (s) Cumulative waiting time: Single backbone 4.16, NeuroSync 9.04 0 3 6 8 11 4.16 Single backbone 9.04 NeuroSync Cumulative waiting time
Cumulative waiting time per turn (s)
Synthetic-session outputs and token use
Metric Single backbone NeuroSync
Tokens per session 30,514 125,933
Sessions with unsupported outputs 54.2% 0%
Full evaluation 8 personas 384 sessions
Evaluation conditions and scopeEight synthetic personas, 384 sessions in total; 144 sessions in the direct comparison with a single-backbone baseline.