Praxis Async

Engineering design document

01 / 08
Part 01Problem framing

The problem is not a lack of clinical data. It is the gap between data and safe action.

Praxis reasons well inside one patient and one session. Value-based care requires it to work across a panel, across time and across channels—without losing clinical context, evidence or human accountability.

Product direction

Praxis connects fragmented clinical data, reasons across patients and time, coordinates the resulting work, and helps clinicians improve—while keeping every recommendation evidence-backed and every clinical action accountable.

Pillar 01

Connect and Understand

Build one reliable, current view from fragmented clinical information.

One-line impact

Clinicians spend time reconstructing the patient before they can reason about care.

AreaWhat breaks

EHR data

Important facts sit across medications, diagnoses, labs, notes and encounters.

Impact: The patient story is incomplete when any one section is reviewed alone.

Claims, labs and pharmacy

Each source arrives with different delays, codes and clinical context.

Impact: New risks can appear late, out of order or disconnected from the chart.

Documents and notes

High-value facts are buried in long, unstructured records.

Impact: Manual chart review is slow and important evidence can be missed.

Messages and inbox events

Symptoms, refill requests and questions change patient state between visits.

Impact: The care plan becomes stale even when the structured chart has not changed.

History, freshness and provenance

Older facts may be superseded, duplicated or contradicted by newer evidence.

Impact: The system cannot safely reason unless it knows what is true, when and why.

Pillar 02

Reason and Prioritize

Decide what matters for one patient—and who matters across the panel.

One-line impact

Attention follows the schedule and the loudest alerts, not always the greatest need.

AreaWhat breaks

Patient-level reasoning

A recommendation must account for the whole patient, not one isolated condition.

Impact: Single-condition advice can conflict with comorbidities, medications or goals.

Care gaps and changing risk

The important question is what changed and whether it requires action now.

Impact: Static reports create noise while meaningful deterioration can remain hidden.

Panel prioritization

A clinician may be accountable for roughly 1,200 patients at once.

Impact: Deep reasoning for everyone is too slow and expensive; simple sorting is too shallow.

Cross-time and condition reasoning

Clinical meaning depends on sequence, trend, conflicts and later evidence.

Impact: A correct fact used at the wrong time can still produce the wrong recommendation.

Practice patterns

Improvement requires comparing decisions and outcomes across patients over time.

Impact: Clinicians cannot see systematic gaps from individual encounters alone.

Pillar 03

Coordinate and Act

Turn a useful insight into owned, trackable and completed work.

One-line impact

A recommendation that never reaches execution does not improve care.

AreaWhat breaks

Inbox work

Messages, results, refills and staff questions require different response paths.

Impact: Clinicians collect context repeatedly before making even small decisions.

Follow-up and outreach

Recommended care must become a task with an owner, timing and next step.

Impact: Patients fall through the gap between identifying a need and contacting them.

Orders and clinical actions

Some work can be prepared automatically; higher-risk actions need approval.

Impact: Without clear autonomy boundaries, the system is either unsafe or not useful.

Care transitions

Admissions, discharges and outside care create time-sensitive follow-through.

Impact: Delay raises the risk of readmission, duplication and medication errors.

Completion tracking

The system must know whether an action was approved, attempted and completed.

Impact: Insight generation becomes another inbox instead of reducing work.

Pillar 04

Trust and Improve

Make asynchronous clinical work explainable, bounded and measurable.

One-line impact

An unwatched error can repeat across a panel before anyone notices.

AreaWhat breaks

Evidence and citations

Every recommendation must connect to patient facts and clinical guidance.

Impact: Without inspectable evidence, clinicians cannot verify the conclusion.

Validation and reconciliation

Agent outputs can conflict, duplicate one another or use unsupported facts.

Impact: The final answer may look confident while hiding unresolved disagreement.

Freshness and accountability

Patient state can change after analysis, and every action needs a human owner.

Impact: A previously reasonable action can become unsafe before it executes.

Evaluation and monitoring

Quality includes omissions, ranking, safety, latency and cost—not just fluency.

Impact: The team cannot know whether a new version is better or merely different.

Feedback, memory and alert burden

The system must learn without silently learning unsafe clinical habits.

Impact: Bad memory creates bias; no memory creates repetitive, low-value alerts.

Transition

These are connected problems, not four separate products.

Better data improves reasoning. Better reasoning creates better work. Safe execution creates feedback that improves the system. The architecture must support that full loop.

Questions & assumptions