Structured intake. Real sources. Deterministic math. Your incidence estimate, delivered to your inbox in minutes — without asking an LLM to guess a number.
AI interprets your text. Deterministic math and real databases calculate the number.
01
Describe the study
Enter the work email, select the audience type, then define core qualifiers and any layered screening criteria.
02
AI parses the text
A language model interprets the audience definition into structured criteria. It does not generate the estimate number.
03
Databases do the math
Structured criteria are looked up against federal probability surveys and provider-approved datasets with deterministic logic.
04
Sourced report by email
A concise report with IR range, feasibility, confidence, and source-level breakdown is delivered to your inbox in minutes.
What the report looks like.
Numbers you can defend, sources you can cite, and a feasibility read you can send to your sample team.
Estimate · v1
Among U.S. adults 30+ with Type 2 Diabetes, nationwide.
Feasible
Incidence rate
Low9.2%
Mid11.4%
High13.1%
Confidence HIGH
Age 30+Census ACS
Type 2 DiabetesNHANES · BRFSS
Nationwide, U.S.Census ACS
Recommendation: Feasibility is strong for standard Rx-adjacent screens. If you need diagnosed and on prescription treatment, expect IR to drop 2–4 points.
Built on sources you would defend to a client.
Five stages. Category-routed evidence. Deterministic math — no LLM-generated numbers.
Deterministic PDF renderer emails the report. Lead logged to Airtable.
The deterministic boundary
Language models are permitted only to (a) parse free text into a controlled taxonomy, (b) classify a criterion into its research category, (c) extract prevalence values from retrieved source documents, and (d) generate the surrounding prose in the final report. Every prevalence, adjustment factor, and final incidence percent is produced by lookup or explicit arithmetic in a Code node. If a figure cannot be sourced at the requested granularity, the criterion is marked UNVERIFIABLE — IR Buddy does not synthesize a number.
Category-routed sources
A short LLM classifier tags each criterion — Demographic, Health, Firmographic, Behavior, or Attitudinal — and selects a hardline domain allowlist from an Airtable-backed source registry before any search runs. Demographic and health criteria pull from census.gov, bls.gov, cdc.gov, nih.gov, and pubmed. Attitudinal criteria go to pewresearch.org, apnorc.org, gallup.com, and ssrs.com. Behavior falls back to Tier 2 syndicated sources like MRI-Simmons.
Tier 1
Federal probability surveys
ACS · CPS · SIPP (Census)
NHANES · BRFSS · NHIS
BLS OES · CES
BEA regional accounts
Tier 2
Commercial & provider-grade
MRI-Simmons · GfK MRI
Scarborough
CDC WONDER · FastStats
March of Dimes registries
Tier 3
Weighted non-probability
Pew Research Center
SSRS · Gallup
AP-NORC Center
Highly-cited survey partnerships
Tier 4
Peer-reviewed & specialty
PubMed · JAMA network
Disease-specialty foundations
Trade & industry publications
Triangulation only — confidence downgraded
Per-criterion retrieval
Criteria fan out through a Split In Batches node so each runs independently. By default each one calls the Perplexity Deep Research API with the category allowlist injected via search_domain_filter — a multi-pass agentic search that reasons over intermediate findings and returns a JSON-schema payload with numbers and source URLs. Criteria that Deep Research cannot resolve (or that need raw PDF-table extraction) fall back to Tavily Advanced + GPT-4o-mini, where Tavily fetches raw content from the approved allowlist and 4o-mini extracts numeric prevalence into the schema. A parse/judge Code node validates every response; a single failure retries with a broadened Tier 2 allowlist, and a second failure marks the criterion UNVERIFIABLE.
Deterministic assembly
Assembly is a pure function of the retrieval outputs. Where a joint distribution exists (e.g., age × gender × ethnicity from ACS PUMS) it is used directly; where only marginals exist, criteria combine multiplicatively with adjustments for recency, published correlation, and panel-reachability. Every report ships a planning interval (low / mid / high) and a categorical confidence tier — HIGH, MEDIUM, LOW, or UNVERIFIABLE — alongside every source URL and score used to build the number.
Deterministic boundary: An LLM parses your text and classifies each criterion. The numbers themselves come from cited sources and documented math — not generation.
To limit token usage during our early demo phase, the retrieval pipeline
is paused. Your inputs were captured but no report has been generated.
Contact us to schedule a live demonstration on your own criteria.