1
Made-up company · real records
The singularity has arrived
Where fiction & reality blur together
Getting my own records was so hard that I made up a company, a website and a product to do it. Wild Hearts Health is not real. The records are.
EVERY CLINIC
Notes
1,241
Lab results
1,154
Visits
1,058
Medications
954
Conditions
155
Wild Hearts Health
2
THE PROBLEM · PDFS SUCK!
PDFs are built for eyes, not agents
My records are spread across 141 providers, and most of them arrive as PDFs.
1
Static
They don’t update as things change.
2
Haphazardly encoded
Huge token consumers to untangle.
3
Made for humans
Designed entirely as a visual representation that works for people. A disaster for machines.
Wild Hearts Health
3
GOING TO THE SOURCE
So I went straight to the source
Wild Hearts pulls structured records through FHIR, the health-data standard that patient portals like MyChart expose. It worked for only 8 of 141 clinics.
FHIR records in the Wild Hearts timeline.
Wild Hearts Health
4
THE HARD PART
Much of the record is still paper
For the other 133 clinics, records requests come back as scanned PDFs: tables, charts and text that a model has to untangle.
Page 2 of a sleep study report.
Wild Hearts Health
5
OCR TEST
Apple Vision vs. Tesseract
Same page, two engines. Characters come through, but rows, columns and units lose their relationships.
Wild Hearts Health
6
LAYOUT-AWARE PARSING
Parsing finds the structure
Reducto marks text, tables and figures with bounding boxes before extracting fields. It’s not private, so and they only offer HIPAA protections for Enterprise accounts.
Wild Hearts Health
7
STRUCTURED OUTPUT
Nonsense into Tables
The parsed result keeps each value with its label and unit, ready for a database instead of a picture.
Wild Hearts Health
8
WHY GROK
Zero data retention
Grok was the only frontier model I could find with zero data retention outside an enterprise contract.
For health records that matters: prompts and outputs aren’t kept after they’re processed, and they aren’t used for training.
The x.ai project for this work is set to ZDR.
47.2 million tokens across 806 requests, all under zero data retention.
Wild Hearts Health
9
SELF-HOSTED AGENT
A private tunnel to Grok
I built my own pipeline and had Grok handle the OCR.
OrbStack
runs Docker on my Mac
↓
Docker
Hermes agent container
↓
x.ai API
private tunnel out
↓
Grok
does the OCR
OrbStack: the Hermes container
Hermes on grok-4.7, checking data retention
Wild Hearts Health
10
GROK'S OUTPUT
The scan, rebuilt as tables
Grok's JSON output becomes clean tables: arousals, body position, oxygen saturation and respiratory measures, each tied to its page.
The same sleep study as slide 4.
Wild Hearts Health
11
WHAT’S NEXT FOR OCR
My own open-weight models
We did Grok the way it likes it: quick & dirty.
Next I’m standing up open-weight models I run myself, on the VPS where wildheartsai.com already lives.
NOW
Grok via x.ai
Fast to set up
no servers to run
Zero data retention
but someone else’s model
General purpose
not tuned for records
NEXT
Open weights on my VPS
Same server
as wildheartsai.com
Post-training
on medical-record data sets
Serious engineering
plus evals to prove it
Wild Hearts Health
12
PRODUCT MAP
Four layers, one record
1
Intake
Records gathering
Automated · manual
OCR → parsing
Convert records into usable information
2
Patient information
Charts
Diagnostics
Imaging · labs · tests
Therapeutics
Procedures · medications · nutrition · supplements · PT / exercise
Visit notes
Visit documentation
Metadata
Longitudinal journey · wearable data · life changes: stress, animals, housing, geography, weather
3
Shared systems
Context management system
Connects the database and charts to each task
Content library
Templates · letters
Wiki
4
Product features
AI chat
Requests / appeals
Exploration / experimentation
Wild Hearts Health
13
FEATURES → CAPABILITIES → CHALLENGES → DECISIONS
Every feature ends in an open question
PRODUCT FEATURE
REQUIRED CAPABILITY
TECHNICAL CHALLENGE
DECISION TO CLARIFY
AI chat
Task-specific patient context. Retrieval with source references.
Select relevant facts within context limits; handle missing or conflicting records; keep answers traceable.
Who uses chat, and for what decisions? What must it decline or escalate?
Gen AI fax
Document drafting + templates. Recipient selection + fax delivery.
Prevent unsupported statements; preserve identifiers; review before sending; track delivery, retries, and duplicates.
Is this for records requests, referrals, or another document? Who approves sending?
Pre-visit
Time-bounded patient summary. Questions + changes since last visit.
Reconcile medications and dates; surface meaningful changes; preserve source links; tailor to the visit.
Is the output for the patient or clinician? What must fit on one page?
Explorations
Patient context + research wiki. Evidence-linked investigation.
Separate patient facts from hypotheses; assess source relevance; handle contradictory evidence and stale research.
What is an exploration’s output: questions, evidence brief, or something else?
Capabilities, challenges and questions are proposals for discussion.
Wild Hearts Health
14
SHARED FOUNDATIONS
What all four features depend on
1
INTAKE
Records intake & extraction
Acquisition → OCR → parsing
Challenges: Source coverage, tables, extraction errors, review queues.
2
MODEL
Patient information model
Facts + events + source documents
Challenges: Dates, units, duplicates, contradictions, provenance.
3
EVIDENCE
Research wiki & controls
Evidence distinct from patient facts
Challenges: Versioning, access, consent, audit trail, retention.
4
CONTEXT
Context management
Choose facts for each task
Challenges: Retrieval, recency, relevance, permissions, context limits.
Wild Hearts Health
15
SHARED SYSTEMS · RECORDS
Engineering an agent-ready medical record
Objective: Build a persistent, structured dataset that agents can query efficiently. Privacy is the primary constraint shaping collection, storage and processing.
CHALLENGE
ENGINEERED RESPONSE AND DETAILS TO CONFIRM
Private processing
Private information must be redacted or processed privately, constraining extraction tools and model choices.
Explored hard-coded processing, self-hosted open-weight models, and provider ZDR/no-training arrangements.
Need: what is deployed versus trialed, and the redaction/validation workflow.
Collection
Fragmented records across 141 providers. Limited connectivity and inconsistent data create gaps, contradictions and errors.
Tried both: automated ETL via Epic/FHIR/HIE (FHIR worked for 8 of 141 clinics) and manual PDF downloads. A blend is likely best; evals will compare FHIR with PDF parsing.
Need: agents/tools, automated steps, and scope of the 30% estimate.
Persistent storage
FHIR access proved too unstable to depend on for continued access to retrieved records.
Built persistent storage on a VPS so retrieved data remains available independently of FHIR access.
Need: database/storage stack, ingestion and retry logic, and the agents’ role.
OCR and parsing
Manual requests yield PDFs with scans, charts, tables and text. Extracting the correct data and relationships is difficult.
Developed OCR and parsing workflows toward structured database records, better organization and lower query-token use.
Need: OCR/models, agent orchestration, and results validated so far.
Still to design: the information structure, content management, and task-specific context management.
Wild Hearts Health
16
SHARED SYSTEMS · RESEARCH
Engineering a multi-agent research wiki
Objective: Build a continuously updated knowledge base combining scientific and community sources, with a judgment layer to separate useful evidence from unreliable claims.
CHALLENGE
ENGINEERED RESPONSE AND DETAILS TO CONFIRM
Source access
Research spans journals, blogs, podcasts, YouTube and private/community platforms. Access differs by source, and Reddit blocks agents in your workflow.
Used Muse for private Facebook groups and Grok for X/Twitter as source-specific collection routes.
Need: the Reddit route and how audio/video sources enter the wiki.
Agent coordination
Different sources require different tools. Research work needs clear responsibilities across several agents.
Set up a central instruction file defining agent jobs. Muse, Grok, Claude and Codex conduct daily research and add to the knowledge base.
Need: each agent’s exact role and handoff rules.
Shared knowledge base
Research from multiple agents needs a common destination that can accumulate knowledge over time.
Built the wiki in GitHub as the shared destination for research contributions.
Need: contribution format, duplicate handling, and the review/merge process.
Credibility and judgment
Scientific publications and community accounts both vary in reliability. Collecting more material does not establish which claims deserve trust.
Identified an adjustable ranking and judgment layer to distinguish useful leads from unreliable claims.
Need: criteria, weights, reviewer responsibility, and what is implemented today.
Open design work: evidence weighting, content management, and selecting relevant research as context for product features.
Wild Hearts Health
17
WHERE IT STANDS
A hard problem,
still in progress
This is a real engineering problem. It couldn’t be finished start to finish in five weeks, and the work continues.
NEXT
Run evals
FHIR vs. PDF parsing
Blend both sources
into one record
Context management
per task
Weight the evidence
research wiki
Wild Hearts Health