Summarization, triage support, EHR integration, and internal ops automation. Human-in-the-loop where clinical judgment matters. $3,999/mo retainer.
- Analyze
- Automate
- Monitor
monthly retainer
Who this is for
You're a healthtech founder or product lead whose market expects AI features in the product. Compliance and evals are non-negotiable. You don't have the runway for an ML hire, and a previous AI attempt stalled because nobody could answer the HIPAA question. You need AI that clinical advisors and investors can actually review.
The pain today
- Competitors are shipping AI features; your product roadmap is stalled on the compliance question
- No ML engineer on the team and no budget to hire one this quarter
- Previous AI spike died in legal review because HIPAA boundaries were not mapped
- Investors and clinical advisors want to see evals and audit trails before approving the feature
- Internal ops — prior auth, intake routing, documentation workflows — still running on manual effort
The outcome you get
- AI features live in your healthtech product on a $3,999/mo retainer
- HIPAA-aware architecture with BAA-covered infrastructure from day one
- Eval infrastructure tracking every prompt change across the feature's lifetime
- Human-in-the-loop gates where clinical judgment cannot be delegated to a model
- Audit logs satisfying regulatory review, investor due diligence, and HIPAA's 6-year retention rule
What AI should — and should not — do in healthtech products
Three patterns deliver value without creating compliance exposure. Summarization: long clinical text, research papers, and patient history condensed for fast clinician review. Triage support: suggesting priority level or care routing from structured inputs, with clinician sign-off before any action. Documentation help: drafting structured notes from captured data for clinician edit and approval. All three keep a human in the decision loop.
There is a hard line. AI should not diagnose autonomously, recommend treatment without clinician oversight, or calculate clinical risk scores that feed directly into care decisions. The distinction between assistance and decision is not just a philosophical point. It is the line that separates a pre-clearance product feature from one that requires FDA review. I build on the assistance side of that line.
AI features live in your healthtech product on a $3,999/mo retainer
Guardrails, evals, and audit trails
Healthtech AI has three non-negotiables before any feature ships. Guardrails: explicit system prompts that define what the model cannot do, output filters that block inappropriate content, and mandatory human review steps wherever a regulation requires it. Evals: golden-set test cases with expected outputs, regression tests on every prompt change, and production sampling with human review to catch drift that static evals miss. Audit trails: every AI interaction logged with inputs, outputs, the user ID, any patient context, and the downstream action taken.
For a regulatory review or investor due diligence, these three are the difference between a defensible AI feature and a compliance fire. I set them up in the first two weeks. They stay in place for the life of the engagement.
2M+: Records processed.
EHR integration and FHIR
Most healthtech AI features need clinical context from an EHR. EHRs that expose FHIR R4 APIs — Athena, Epic via App Orchard, Cerner, DrChrono — integrate for reading patient context into the AI layer. The AI reads that data, produces output, and writes back to the EHR as structured or unstructured notes with clinician approval.
Full EHR write integration requires careful testing. EHRs do not forgive sloppy writes, and write errors can surface in patient records. The realistic integration timeline is 4 to 8 weeks on top of standard AI feature work, depending on the EHR vendor and whether SMART on FHIR OAuth flows are already in place. Read-only integrations for summarization or triage context are faster — typically 2 to 3 weeks.
For healthtech products that need to pass hospital procurement, the EHR integration is often what gets you past the review. I build to the EHR's published standards, not workarounds.
Internal ops automation for healthtech teams
AI inside the product is half the picture. Healthtech teams run manual processes that AI handles cleanly: prior authorization prep (structuring clinical data for payer submissions), intake routing (triaging inbound patient requests to the right care pathway), documentation workflows (converting call transcripts or form submissions into structured records for clinical review), and data reconciliation across systems.
These internal automations do not require FDA review. They reduce administrative burden on clinical staff and accelerate revenue cycle timelines. The Reevia engagement I ran processed 2M+ records across four systems into a single source of truth, with sub-50-second sync — the same data-pipeline discipline applies to healthtech internal ops. The difference is that healthtech data pipelines have PHI boundaries that require explicit architecture.
Model choice: hosted vs. self-hosted
Hosted options with HIPAA BAAs: Azure OpenAI, AWS Bedrock with Claude, Anthropic enterprise contracts. These are easier to operate, faster to iterate on, and cover the majority of healthtech AI use cases. Self-hosted open-source — Llama 3/4 via vLLM, Mistral — is the right choice when clinical trial data or particularly sensitive PHI cannot leave your infrastructure, or when per-call cost at scale makes hosted options uneconomical.
Self-hosted adds real operational overhead. Somebody owns the model runtime, the update cycle, and the latency budget. I help make this call in week one based on actual data flow mapping, not guesswork. For most healthtech startups at seed to Series B, hosted with proper BAAs is the correct answer. Self-hosted becomes worth it at higher volume or when a specific compliance requirement mandates on-prem processing.
Performance and latency in clinical workflows
Clinicians do not tolerate slow AI. The targets I hold to: under 3 seconds for inline AI assistance, under 10 seconds for full patient history summarization. For processes that need longer, streaming output so clinicians see partial results immediately.
At Cuez I drove API response time from 3 seconds to 300 milliseconds — 10x faster — under production load. That same discipline applies to AI features. Model routing keeps latency under control: a smaller, faster model handles straightforward tasks; a larger model takes the complex cases that warrant more processing time. Repeated queries get cached. The performance architecture is designed before the first prompt goes to production, not bolted on after complaints start.
Pricing and what the retainer covers
The AI Automation retainer is $3,999/mo. It covers AI feature integration, HIPAA-aware architecture, eval infrastructure setup and maintenance, model monitoring, and iteration as the product evolves. The 14-day money-back guarantee applies. Cancel anytime. All code is yours under Work Made for Hire. NDA and BAA standard.
LLM and infrastructure costs pass through at cost — typically $500 to $3,999/month depending on model choice, call volume, and whether you are on hosted or self-hosted infrastructure. For healthtech companies also building out the core product, the AI Automation retainer often runs alongside an Applications subscription. The two engagements share cadence and context, which reduces the coordination overhead that comes from splitting product and AI work across different contractors.
Recent proof
A comparable engagement, delivered and documented.
Four systems, one source of truth: HubSpot visibility for one of Brazil's largest vet networks
Built a custom integration layer for Reevia that connects four source systems into HubSpot for one of Brazil's largest veterinary companies. Over 2 million records processed with full normalization. Any lead from any system is inside HubSpot in under 50 seconds, standardized and ready to use.
Read the case studyFrequently asked questions
The questions prospects ask before they book.
It means PHI flows only through infrastructure covered by a signed Business Associate Agreement — Azure OpenAI, AWS Bedrock, or Anthropic enterprise. Clear code-level boundaries separate PHI-touching calls from general-purpose AI calls. Role-based access control limits who can query the AI layer. Audit logs capture every input, output, user, and patient context. Zero-data-retention tiers are configured where prompts must not persist. The architecture decisions are documented for your compliance and legal team to review.
Golden-set evals: 50 to 200 hand-crafted test cases with expected outputs for each feature, run automatically on every prompt change. Production sampling with human review catches distribution shift that static evals miss. For clinician-facing outputs, 100 percent human review gates releases until you have enough production confidence to relax the threshold. Eval tooling choices — LangSmith, Braintrust, or a custom harness — depend on your scale and team preferences. Evals are non-negotiable; AI features without them drift silently.
Yes, for EHRs with FHIR R4 APIs — Athena, Epic via App Orchard, Cerner, DrChrono. The AI reads clinical context from the EHR, produces output, and writes back as structured or unstructured notes with clinician approval. Read-only integrations for summarization or triage typically take 2 to 3 weeks. Full read-write integrations take 4 to 8 weeks on top of core AI feature work, depending on EHR vendor and whether SMART on FHIR OAuth is already configured.
For BAA-covered hosted providers, prompts and outputs are retained per the provider's BAA terms — typically 30 to 90 days for abuse monitoring, with opt-out for longer retention. Where retention creates risk, I configure no-retention API tiers or route those calls to self-hosted models. Audit logs are kept for 6 years per HIPAA requirements. Patient deletion rights cascade through the application layer — deleting a patient record triggers cleanup of any AI-processed data tied to that patient.
Pre-clearance features keep humans in the decision loop: summarization of clinical notes for clinician review, triage support that suggests routing but requires clinician approval, documentation drafts from structured data for clinician edit. These do not constitute diagnostic aids or clinical decision support tools under FDA's Software as a Medical Device guidance when a human reviews and approves every output. Features that make autonomous clinical decisions — diagnostic aids, risk scoring that directly drives care — require FDA clearance. I build on the pre-clearance side.
Prior authorization prep is the clearest win — structuring clinical data for payer submissions reduces the manual time clinical and admin staff spend per case. Intake routing, documentation conversion from transcripts or forms, and cross-system data reconciliation all automate cleanly. These are internal workflows, not patient-facing AI features, so they fall outside FDA Software as a Medical Device scope. PHI boundaries still apply, but the compliance overhead is lower than for product-facing clinical AI.
The AI Automation retainer at $3,999/mo runs independently or alongside an Applications subscription at $4,999/mo. When both are active, the AI and product work share a single engagement context — same codebase, same architecture decisions, same daily async cadence. That removes the coordination cost of splitting AI and product work across separate contractors. LLM and infrastructure costs pass through at cost on top of the retainer, typically $500 to $3,999/month depending on model and volume.