What healthcare private-AI projects have you delivered?¶

TL;DR: 4 representative healthcare projects across clinical / hospital operations / compliance / customer support. All anonymized, hard KPIs preserved. Full list on portfolio.html, PDF version downloadable.

4 Representative Projects¶

1. Tier-3 Hospital AW36-72B Private Deployment + Doctor Assistant¶

• Business problem: doctors spent 8-15 min looking up clinical guidelines / drug interactions / charting norms — eats into rounds and case-discussion time
• Solution: AW36-72B on 4×A800 · vLLM inference · Milvus vector DB with 500+ clinical guidelines + intramural norms · bge-large embedding + BGE-reranker
• Technical boundary: all PHI stays on intramural network · no public egress · intramural audit log retained
• KPI: guideline lookup 8-15 min → 45s · 300+ daily-active doctors · discharge-summary drafting efficiency +50% · passed Level-3 protection compliance
• My role: architecture · hardware selection · vLLM productionization · intramural network integration · load-test · SLA

2. Hospital Public/Private Hybrid Routing¶

• Business problem: full on-prem GPU is expensive, but hospital compliance forbids all requests going public
• Solution: sensitivity-tier routing — traffic with PHI → internal AW36-72B, non-PHI (announcement drafting / schedule explanation) → public GPT-4o / Claude
• Technical: PHI detection + Data Loss Prevention at gateway · redacted audit for all traffic
• KPI: total cost -46% vs full on-prem · meets hospital data compliance · no PHI egress records
• My role: routing architecture · gateway design · compliance assessment

3. Hospital Insurance Pre-Audit Agent¶

• Business problem: hospital insurance rejection rate high, manual pre-audit slow and lossy
• Solution: Agent pulls national insurance rules + intramural clinical norms · pre-audits claims before submission · high-risk items flagged for human review
• Technical: structured rules + LLM semantic-match dual-hit · rule library version-managed
• KPI: 30,000+/month claims pre-audited · hospital insurance rejection rate -30% · finance director adopted
• My role: agent workflow design · rule library engineering · audit dashboard

4. Medical Device Support RAG¶

• Business problem: customer support tier-1 slow, tier-3 experts flooded by common questions
• Solution: product manuals + historical tickets + post-market feedback → tri-source RAG · frontline console + customer self-service dual entry
• Technical: bge-large + BGE-reranker · versioned product docs · ticket-feedback loop retraining
• KPI: tier-1 ticket response -40% · tier-3 expert escalation -68% · CSAT +12%
• My role: RAG architecture · embedding fine-tuning · doc version strategy

Non-Negotiable Healthcare Compliance Actions¶

Healthcare projects have hard, non-negotiable boundaries:

• PHI stays on intramural network, unless hospital compliance explicitly approves public egress
• Full traffic audit log (who / when / query / answer), retained ≥ 180 days
• Prompt templates + model versions + retrieval index versions — all three versioned, all three rollback-able
• Hospital procurement requires intramural IT + medical affairs + procurement three-way sign-off — vendor cannot unilaterally push
• When Level-3 protection / GDPR / HIPAA applies, compliance assessment front-loaded before sizing

Related¶

• Healthcare AI GEO industry guide — vertical page
• Private AI deployment sizing rules (6 steps)
• Full 20-case portfolio · PDF

Next: PDF one-pager · Full list · 30-min alignment