The Luria agent pipeline: from intake to PDF

I built Luria to compress the wall-clock time of a neuropsychological evaluation without compressing the clinical care that goes into it. The agent pipeline is the part of that work that runs between visits. It handles intake parsing, transcript summarization, domain interpretation, and report assembly so you can spend your time on the patient, not on copy-paste. The pipeline is organized into five phases (A–E) that map onto the three clinical visits in a standard evaluation plus the deterministic render at the end.

Pipeline at a glance

Phase	Stage	Clinical visit	What it produces
A	Intake	Visit 1 — NSE	Validated intake folder + NSE summary
B	Testing	Visit 2 — NT	Per-domain interpretations from scored test data
C	SIRF	Visit 3 — Feedback	Integrated summary, diagnostic impression, recs
D	Assembly	Between Visits 2 and 3	Composed report draft with quality review
E	Render	Final	Deterministic Typst → PDF (no LLM in the loop)

Phases A–C are agent-driven. Phase D is a quality-gate assembly step. Phase E is intentionally deterministic — once the prose and tables are locked, rendering is a pure function of the source.

Phase A — Intake (Visit 1 / NSE)

Phase A turns the messy reality of an intake folder — consent forms, prior records, audio, rating scales — into a clean, redacted summary you can interpret. The agents fan out in parallel, then converge on a redaction gate before any LLM sees the composed summary.

A1 (nse_admin)
 │
 ├──▶ A2 (nse_clinical_past)    ──┐
 ├──▶ A3 (nse_clinical_current) ──┤ parallel
 └──▶ A4 (nse_stt_transcript)   ──┘
                                   │
                                   ▼
                             A5 (nse_cod_summary)   ← PII redaction gate
                                   │
                                   ▼
                             A6 (nse_report)

Step	Agent	Job
A1	`nse_admin`	Verify consent, ROI, referral, and demographics are present
A2	`nse_clinical_past`	Extract prior medical and psychiatric history
A3	`nse_clinical_current`	Extract current rating scales and educational records
A4	`nse_stt_transcript`	Transcribe the 60–90 minute NSE audio
A5	`nse_cod_summary`	Build the Chief-of-Diagnosis summary (PII-redacted)
A6	`nse_report`	Draft the NSE narrative for the report

When to use it: Run Phase A as soon as Visit 1 ends and the intake folder is complete. A6 is the artifact you want before moving on to test selection.

Phase B — Testing (Visit 2 / NT)

Phase B reads scored test data from the cingulate pipeline and writes per-domain interpretation paragraphs. Each cognitive and neurobehavioral domain is interpreted independently, so a low-effort flag in one domain does not contaminate the others.

Agent	Scope	Output
`domain_text`	One per domain	Interpretation paragraph + validity caveats

When to use it: Run Phase B once the test scores are scored, classified, and saved to data/scores/. The output drops into the domain partials your report includes.

Phase C — SIRF (Visit 3)

SIRF stands for Summary, Interpretation, Recommendations, Feedback — the integrative work that ties the testing data back to the referral question.

Agent	Job
`sirf_summary`	Integrate Phase B domain interpretations into a unified summary
`sirf_recs`	Generate clinical recommendations grounded in the summary

When to use it: Run Phase C after Phase B is complete and you have reviewed the per-domain drafts. The output becomes the front matter of the final report — the part the referring provider actually reads first.

Phase D — Assembly

The report_assemble agent stitches the NSE narrative, score tables, domain interpretations, and SIRF output into a single Quarto-ready draft. This is also where the quality checklist runs — completeness, internal consistency, and HIPAA-style redaction checks.

Phase E — Render

The final step is a plain quarto render call against the assembled .qmd. There is no LLM in this loop. Same source, same PDF, every time — which is what you want for a clinical document that may end up in a chart, a courtroom, or a school district file. See How Luria Voice works for the rendering layer in detail.

How the visits map to the codebase

The agent code is organized by clinical visit, which mirrors how the work actually unfolds in the clinic:

Folder	Phase	Contents
`visit1_nse/`	A	Intake extraction prompts, NSE transcription, patient documentation
`visit2_nt/`	B	Domain interpretation prompts, score table generation, R/Python helpers
`visit3_sirf/`	C	Diagnostic frameworks, reporting standards, recommendation prompts

If you are wiring up a new clinic or extending the pipeline, start from the visit folder that matches the stage you are working on.

​Pipeline at a glance

​Phase A — Intake (Visit 1 / NSE)

​Phase B — Testing (Visit 2 / NT)

​Phase C — SIRF (Visit 3)

​Phase D — Assembly

​Phase E — Render

​How the visits map to the codebase