PRD & Architecture

Conversational Qual. Real clinical reasoning at scale.

Sagan brings AI-moderated async In-Depth Interviews into pharma research so a brand team gets 12 oncologist conversations in 48 hours, not 8 weeks - at a quarter of the agency cost.

Live Moderator· 1L EGFR+ NSCLC

Preview loop

Coverage0%

Patient profile

Regimen chosen

Decision driver

CNS consideration

Alternatives weighed

Switch triggers

Probe rate0.00 / ≥0.60

This loop is illustrative. The real moderator runs a live AI interview.Run a real interview

3×
Target depth: 48h
Target turnaround: ¼
Target cost

02 · The Problem

Pharma qual produces decks, not decisions.

The bottleneck isn't the moderator's skill - it's the calendar. Every step is sequential, so the deck lands after the brand question has already moved. Async AI moderation collapses the sequence.

Agency IDISagan

Timeline6-8 weeks48-72 hours

Cost (n≈20)$150K-$300K~¼ the cost

ExecutionSequential, human-bottleneckedAsync, runs in parallel

SynthesisAt study endFires on each session close

ArtifactA static deck (arrives stale)A live dashboard

03 · Users and Buyers

Four people, one product.

BUYER

Priya - Insights Director

Top-20 pharma

Job to be done: Get directional read on physician sentiment 6× faster than current agency cadence.
Top frustration: Spends 6 weeks waiting on a deck that arrives stale.
What success looks like: Decision-ready findings in <72h with verbatim depth.

CONSUMER

Marcus - Brand Manager

Lung cancer franchise

Job to be done: Understand why oncologists pick a regimen so I can sharpen messaging.
Top frustration: Insights team's deck doesn't answer my actual targeting questions.
What success looks like: Ask follow-up questions in plain English and get cited quotes back.

RESPONDENT

Dr. Patel - HCP Respondent

Academic oncology

Job to be done: Share clinical thinking without burning 90 minutes on a video call.
Top frustration: Email recruitment + Zoom scheduling = wasted week, no honorarium worth it.
What success looks like: 20 minutes async, peer-to-peer feel, real clinical depth captured.

INTERNAL

Ankur - ZoomRx BD Lead

Internal commercial

Job to be done: Sell paid pilots faster by showing a live working demo, not slides.
Top frustration: Prospect sees pitch, asks to try it - and nothing exists to try.
What success looks like: Send a link, prospect runs a 5-min interview themselves, signs.

04 · Product Architecture

Deterministic where it matters. Probabilistic where it helps.

Deterministic · rule-based codeProbabilistic · LLM

Full component / model / determinism table

Component	Output	Model	Determinism
Discussion Guide (typed)	Section prompts + must-capture items	-	Deterministic · lib/sagan/discussionGuide.ts
Guide retrieval	Keyword-scored probe + branch selection (BM25-lite, not vector RAG)	-	Deterministic · engine.ts selectProbes
Moderator Prompt Builder	System prompt per turn, fed the live gaps	-	Deterministic · moderatorPrompt.ts
OpenRouter LLM	Moderator turn	claude-sonnet-4.5	Probabilistic
Validation + Branch Engine	Coverage-gated section advancement + branch triggers	-	Deterministic · engine.ts
Quality metrics	Probe rate / depth / coverage / brand-leakage, per transcript	-	Deterministic · engine.ts computeQualityMetrics
PV / AE listener	EphMRA 4-criteria adverse-event flags (patient / reporter / product / event) + seriousness	-	Deterministic · pvListener.ts
Supabase persistence	Sessions + messages + synthesis	-	Deterministic
Synthesis Pipeline	Themes / drivers / brand perceptions / archetypes	claude-sonnet-4.5 (JSON mode)	Probabilistic w/ schema
Insights Dashboard	Charts + NL query + transcript drill-in	-	Deterministic

Three deterministic guarantees · implemented in lib/sagan/engine.ts

Mandatory section coverage - deriveSectionState extracts must-capture items from the respondent's words and blocks the transition until they are captured (an anti-stall cap of 2x target turns is the only turn-based guard). Watch the validator chips fill in the Live Moderator.
Structured synthesis output schema - JSON-mode forces themes/drivers/brand/archetypes into typed shapes.
Falsifiable quality SLAs - computeQualityMetrics scores probe rate, depth, coverage and brand-name leakage from each transcript, so a shallow interview fails on-screen rather than passing by assertion.

05 · Conversation Flow

Six beats. Then it synthesizes itself.

~20 minutes async. The last two nodes (accented) fire with no human in the loop - the same momentum you watch in the live progress bar.

01Warm-up~2 turns, persona context
02Patient walkthroughrecent case, branching
03Decision driverswhy this regimen
04Brand perceptionnamed brands, unaided
05Stimulus reactmessage / claim test
06Wrapfinal probe + close
07Synthesis firesthemes / drivers / archetypes
08Dashboard updatesnew completed session live

06 · Discussion Guide Framework

Six sections. Twenty-two minutes.

Warm-up

2 turns

~4 min

Must capture

practice setting · avg new EGFR+ patients per quarter

Recent Patient Walkthrough

4 turns

~8 min

Must capture

patient profile · chosen regimen · reason for choice

Decision Drivers

3 turns

~6 min

Must capture

top 3 drivers · CNS handling preference · toxicity threshold

Regimen Perceptions

3 turns

~6 min

Must capture

perceived strength and concern for each regimen

Stimulus React

2 turns

~4 min

Must capture

shift trigger threshold

Wrap

1 turns

~2 min

Must capture

one fix the respondent would make

07 · Quality Metrics

Three numbers that decide go / no-go.

All three are computed deterministically from the transcript by engine.ts and shown on the Live Moderator's summary scorecard - so a weak interview fails them on screen instead of passing by assertion.

Probe Rate

Computed

Probes per substantive answer, counted from moderator turns that dig vs. open a topic.

Target ≥ 0.6

Computed live in the Live Moderator (engine.ts). Measures digging, not just running the guide.

Depth Score

Computed

0-5 specificity of each respondent answer, scored on length, numbers, clinical markers and reasoning.

Target avg ≥ 3.5

Computed live per transcript. A generic interview scores low and visibly misses the SLA.

Brand-name integrity

Computed

Off-guide brand names (Tagrisso, Rybrevant, Lazcluze...) the moderator introduced first.

Target = 0

Computed live. Replaces the unfalsifiable '0 hallucinations' claim with a checkable rule break.

08 · Competitive Positioning

Pharma-native depth. No horizontal tradeoffs.

Sagan / ZoomRx

Verified pharma panel + clinical-context guide + regulatory rigor. The wedge competitors can't copy without rebuilding the panel.

Horizontal AI moderation

Outset, Strella, Listen Labs. Deep AI moderation but no pharma panel - every study needs custom recruitment.

09 · Roadmap

Twelve months, three horizons.

90 days

Land 3 paid pilots with Top-20 pharma BD relationships
Ship the EGFR+ NSCLC vertical end-to-end (live moderator + dashboard)
Probe Rate ≥ 0.6 across 50+ live interviews

180 days

Voice-first mobile capture for the respondent
Branching templates for 5 more oncology indications
Cross-session NL query graduates from beta

12 months

Expand from HCP qual to payer + patient async qual
Real-time client co-watch (insights director sees responses live)
API endpoint so client research teams can embed Sagan in their own workflows

10 · Provenance

Every claim is labelled.

What is real vs illustrative

Real code

Coverage-gated validator, branch engine, guide retrieval (engine.ts), the EphMRA 4-criteria AE listener (pvListener.ts), a live claude-sonnet-4.5 call, computed quality metrics and Supabase aggregation - each with a labelled deterministic fallback if no API key is set.

Public-sourced

ZoomRx panel scale (60,000+ HCPs), coverage of 19 of the top 20 pharma, and competitor raises - all from public disclosures.

Illustrative

Pricing, turnaround and cost targets, seeded dashboard figures (badged in-app), the sample synthesis shown when live AI is off, and PV routing after an AE flag (detection is real; the 24h sponsor-inbox handoff is not performed in a demo).