Get the free AI Delivery Playbook
AkveoContact us

A HIPAA-Compliant Platform for Opioid Crisis Response Teams

A healthcare platform that helps post-overdose response teams find people nobody referred, hand them off across independent agencies and prove the handoff worked, online or offline.

A HIPAA-Compliant Platform for Opioid Crisis Response Teams

[ Client ]

About the Client

Goldie Health is a digital health platform created to support agencies identifying and engaging individuals at risk of opioid overdose. The platform helps response teams deliver personalized care by integrating with electronic patient care reports (ePCRs), electronic health records (EHRs), and local community referral networks. It addresses barriers to treatment by factoring in social determinants of health such as housing instability and financial constraints.

87%
Fewer overdoses in the pilot county, year one
72%
More patients identified than EMS referrals
~2 hrs
Admin time saved per staff member a day

[ Challenge ]

Overdose Response Ran on Spreadsheets, Phone Calls and Memory

Work lived on paper. Each county kept its own spreadsheets, so nothing could be compared and no program could be evaluated beyond a single jurisdiction.

A person existed only if someone filed a referral. If EMS or the hospital didn't refer them, the team never heard of them, and nobody knew how many people were lost that way.

Handoffs between agencies disappeared. A referral to a clinic or a housing program was a phone call and a note in a notebook. "Did they get there?" had no answer, and it is the only question that matters.

The field had no system at all. Outreach happens in cars, shelters, tent encampments and rural areas without coverage. Anything that needed the internet became a second job done in the evening from memory.

The regulatory bar is unusually high. Substance use data falls under both HIPAA and 42 CFR Part 2, which is stricter on disclosure. Several independent agencies have to work with the same person, and none can see more than it is entitled to.

[ Solution ]

Building the Solution

Akveo built three apps: a mobile app for field work, one web app whose functionality changes with the user's role, and a super-admin console for managing all counties. Behind them sit a pipeline that pulls data from EMS systems and a public API for integrations.

AI detection across every EMS record

Every EMS call record now passes through the system, not only the cases someone referred. An AI service flags records that look like a substance-related episode. It works only on de-identified data and returns a decision with a confidence level, reasoning and the key indicators, so a person can disagree. Structured output keeps the classification reproducible, and a rule-based fallback takes over when the model is unavailable.

Each county has its own rule profile with hard yes and hard no conditions, calibrated with county clinical specialists on live data. It separates the real edge cases, such as choking with no other signs, children, alcohol without a named substance, overdoses of prescribed medication and household chemicals. One kind of error wastes a team's trip; the other costs a person.

Data arrives through four channels, because nobody has a clean API: ESO and ImageTrend, the two dominant US EMS charting platforms; HL7 v2 over MLLP with mTLS and deduplication; encrypted secure email; and web forms with per-jurisdiction handlers.

We've moved beyond being a workflow management tool into a lifesaving detection platform.
Chris Martin, CEO, Goldie Health

Cross-agency referral workflow

A referral is not a message but a request with an explicit lifecycle that both sides see: not sent, sent, received, accepted, in progress, completed, or declined, not received and cancelled. The sender is notified at every step, so nobody calls around for a status.

Two independent organizations edit the same referral with different rights on the same fields. The right to a field depends on who you are relative to the record, what state it is in and what type it is. The logic lives in a dedicated engine backed by a documented state matrix the team tests against, instead of conditions scattered across the interface.

Inbound referrals list with the referral details panel and accept and decline actions

Offline-first mobile app with version-safe sync

A visit to a basement, a shelter or a car with no signal still gets recorded. The phone is the real record, not a draft until you're back at the office. An encounter can be saved unfinished, with only its type and capture method required.

Devices can stay offline for weeks while field staff run different app versions. Offline changes are stored as intent, a command tagged with its schema version, and upgraded to the current model when they reach the server. In the other direction, each device gets data in the format its version understands. The product kept changing, and nobody in the field had to update just to sync.

Sync downloads only what changed. The server keeps a separate clock for each family of entities, and IDs are generated on the device, so offline records survive the round trip and can reference each other.

The patient card has an AI summary that streams section by section over server-sent events and is never stored.

Goldie mobile app showing a patient's tasks and encounters

Context-based access control

Nobody configures access by hand. You see a patient because you are on their team, have a task for them or are running a visit; when the work ends, access goes away. Each grant stores the source that created it, so it can be revoked automatically and explained later. Access changes travel with the data: a device learns what it may no longer see and deletes it.

For a patient who isn't assigned to the worker but needs help now, a county can enable a controlled path. The worker finds the person only by full last name, sees a strictly limited set of fields to confirm identity, and gets access only after confirming care was provided. The grant is recorded in the patient's history as a time window.

There are six roles, each with its own access to sections, and that access differs between the mobile app and the web.

Goldie patient list with organization, program, status, assigned staff and condition for each patient

Multi-tenant isolation with row-level security

All counties share one schema, and row-level security in PostgreSQL limits every statement to the current tenant. If no tenant is set, nothing matches, so a bug returns an empty result instead of another county's data. Cross-tenant administration lives behind a separate API with its own role. A new county is provisioned in one operation that rolls back every step, in AWS and in the database, if any step fails.

De-identified analytics layer

Dashboards for counties and funders run on a separate de-identified schema. Database triggers remove or mask all 18 HIPAA Safe Harbor identifiers in real time. Dates are cut to month and year, and patient IDs become one-way hashes, so one person's path is still traceable over time without knowing who they are. Analytics for one county cannot show its neighbor's data.

Real-time chat and patient SMS

All communication lives in one section: SMS to patients, internal patient discussions, channels and direct messages, each with an unread counter. SMS is kept apart visually and technically, so it can't be confused with messages to colleagues.

Each county has its own number. A patient conversation is keyed by phone number rather than by patient, because one phone often belongs to several people. Chat runs over a live connection, only for the conversations the user is allowed to see, with push notifications when the app is closed.

Goldie mobile app showing SMS to patients, patient chats and a patient profile

A domain model for clinical and social services

The platform covers more than 30 service types, most of them non-clinical: food, phones, clothing, shelter, long-term housing, utility bills, documents, insurance, legal help, employment and transport. The clinical side includes addiction treatment, harm reduction, syringe exchange, wound care, HIV and hepatitis C testing and psychiatry.

Clinical scenarios are built in, not bolted on: buprenorphine induction with the COWS scale, and assessments on a calendar from the initial one through 2, 3, 7 and 14 days to 1, 2, 3, 6 and 9 months.

[ Results ]

Impact that Matters

Over 2.5 years, a team of up to 11 Akveo specialists, including four full-stack engineers, two mobile engineers, a UI/UX designer, two business analysts, a QA engineer and a DevOps engineer, worked with the Goldie Health team to reach these results.

Fewer Overdoses and Deaths

In the year the pilot county launched its response team on Goldie, overdoses fell 87% and drug-related deaths fell 46%.

More People Found

Goldie identified 72% more patients than EMS referrals and 300% more than hospital referrals.

Designed for HIPAA and 42 CFR Part 2

Substance use data falls under both HIPAA and the stricter 42 CFR Part 2. Access follows the work, counties are isolated in the database, and analytics run on data stripped of all 18 Safe Harbor identifiers.

[ Tech Stack ]

Technologies Behind the Project

Web

ReactTypeScriptTailwind CSS

Mobile

React NativeExpoTypeScriptMobXRealmExpo RouterReact Hook FormYup

Backend

Node.jsNestJSTypeScriptKnexPostgreSQLSocket.IOSendbirdAWS (Cognito, S3, SNS, Pinpoint)

AI & Data Ingestion

OpenAI structured outputServer-sent eventsNxHL7 v2 / MLLPRedisESOImageTrend

Analytics

Drizzle ORMApache SupersetDocker

Visuals

Goldie web dashboard with active cases, patient status, referral activity, a priority queue and an AI assistant panel

[ Case Studies ]

Explore More Success Stories

We have helped over 200 businesses grow their value and improve how they work through better software.

Ready to Build with AI?

Let’s discuss your goals and how AI can help you achieve them faster.

Dmitry Klim
Dmitry Klim
Head of Growth
Clutch
5900 Balcones Drive #21729, Austin, TX 78731[email protected]+1 (512) 921-9631