How It Works

On the surface, it's what you'd expect.

Engine

Report

That "engine" step is not one step.

Every product is built from modules — each one a full adversarial pipeline where multiple AI models draft, review, and attack each other's work. A single report can run 8, 19, even 25 modules before you see it.

Example: Entity Intelligence — 19 modules

Company Overview

8 stages

Financial Analysis

8 stages

Competitive Landscape

8 stages

Market Position

8 stages

Leadership & Org

8 stages

Risk Assessment

8 stages

AI Exposure

8 stages

Regulatory

8 stages

Customer & Revenue

8 stages

Technology Stack

8 stages

Supply Chain

8 stages

Growth Vectors

8 stages

M&A History

8 stages

ESG & Social

8 stages

Valuation Signals

8 stages

IP & Patents

8 stages

Partnerships

8 stages

Talent & Culture

8 stages

Cross-Module Synthesis

8 stages

↑ Let's open one up

Inside One Cortix Module

Input

Decompose

Organize

Draft

Review

Attack

Finalize

Report

Decompose

10–20 search strings generated

Your query and any uploaded documents are analyzed and decomposed. The engine creates distinct search strings designed to cover the various topics and angles required — market landscape, financial analysis, regulatory, AI exposure, etc.

↓

Hundreds of results per module

Search strings are sent out to multiple search engines simultaneously — casting a wide net across sources in parallel.

↓

Organize Results

Deduplicated, scored, ranked

Results are deduplicated, scored for relevance and authority, then ranked. Math required to optimize results quality — tokens for depth of each result, trade-off with number of results to review from ranked results.

↓

Analyze & Draft

Preliminary report constructed

Reservoir of content is analyzed and a draft report is constructed. Tone and strength of assertions (harshness of critique) can be dialed up or down.

↓

Independent Review

Different AI model reviews independently

A separate AI model independently reviews the draft — verifying claims, checking reasoning, and conducting its own searches to confirm or challenge findings.

↓

Adversarial Attack

Third AI from a competing company

Attacks the draft — finding gaps, weak claims, and counter-evidence. It's not collaborating, it's challenging.

↓

Finalize

Only surviving claims make the cut

All adversarial challenges are incorporated into a hardened final version. Only claims that survive all perspectives make the report.

↓

Score & Deliver

Formatted & delivered

Report formatted based on direction — free form based on content created, or strict rules.

Inside One Cortix Module

Build

Input

Decompose

10–20 search strings

Query and documents analyzed, decomposed into distinct search strings covering all required angles.

Hundreds of results

Strings sent to multiple search engines simultaneously — wide net cast across sources in parallel.

Organize

Scored & ranked

Results deduplicated, scored for relevance and authority, then ranked. Optimized for depth vs. breadth.

Draft

Preliminary report

Content analyzed and draft report constructed. Tone and assertion strength can be dialed up or down.

Challenge & Deliver

Review

Different AI model

A separate AI independently reviews the draft — verifying claims, checking reasoning, running its own searches.

Attack

Third AI, competing company

Attacks the draft — finding gaps, weak claims, counter-evidence. Not collaborating, challenging.

Finalize

Only survivors make the cut

All adversarial challenges incorporated. Only claims that survive every perspective make the final report.

Report

What We Use

Search Engines

Brave Search (Parallel)

Claude Search

Perplexity

Gemini Search

Grok Search

AI LLM Model Providers

Anthropic — Claude family

xAI — Grok family

Google — Gemini family

OpenAI — GPT family

Various models within each provider can be selected depending on the task — speed, depth, cost.

OpenAI is not currently in use for reasoning due to rate limits that create reliability problems at production scale. It can be swapped back in at any time.

LLM Agnostic with Architecture Flexibility

Everything is interchangeable. As models get better or worse, we swap them in or out across all products at the base level. No product is locked to a single model or provider. The engine is built to be model-agnostic — what matters is the methodology, not any single vendor.

Why This Is Unique

Cortix Stack

No Ceiling

Revenue Quality Module

Competitors Analysis Module

Regulatory Dynamics Module

AI Exposure Module

Portfolio Concentration Module

Event Dependencies Module

Potential Buyers Module

Bull & Bear Case Module

Cortix Engine

Every module can be stacked on top of each other — compounding depth across any number of dimensions
Reports run from 15 minutes to hours to days, depending on building blocks used
Add modules, remove modules, reconfigure per use case
The architecture scales with the research purpose, not against it

Cortix Lens

Best Idea Wins

Draft

Challenge

Verify

✓

Hardened

Three providers compete · Only surviving claims make the report

Most tools produce summaries — neutral recaps of what was found. We produce assertions. Every module is instructed to take a position, cite the evidence, and surface the tension. The system has opinions. And it defends them.

Not this

"Revenue grew 12%."

This

"Revenue grew 12% while the core business contracted — growth is acquisition-driven and may not be sustainable."

Assertion, backed by evidence, with the tension built in.

Assertion · Evidence · Tension

Assertion A clear analytical position. Not a fact dump — a judgment that takes a side and says what it means.

Evidence Sourced, dated, linked. Every claim traces back to what was actually found — not what the model "knows."

Tension What complicates this? What contradicts it? The system doesn't hide uncertainty — it structures it. Required on every assertion.

Confidence Auto-capped by evidence quality. Thin evidence caps at medium. No exceptions.

Summaries tell you what happened. Assertions tell you what it means.

Patent pending

Cortix Studio

Cursor Meets Palantir

Describe

natural language

build

● Live

your product

Natural language wrapper on top of the engine
Describe what you want to research, how you want it structured — Studio builds it
No code, no configuration, just intent
Build your own re-usable research product in minutes. Build product variant to match your use case

studio

Describe your product

I need a healthcare staffing analysis with regulatory review, market sizing, and competitor mapping across 3 regions...

Build

init_module()

├ market_sizing

├ regulatory

├ competitors

├ staffing_trends

├ region_east

├ region_west

└ region_central

modules: 7 ✓

config_search()

brave: connected

claude: connected

perplexity: ready

gemini: connected

grok: connected

dedup: enabled

ranking: active

search: 5 ✓

set_reasoning()

anthropic: opus

xai: grok-3

google: gemini

adversarial: on

tension: required

confidence: auto

tone: balanced

llms: 3 ✓

build_product()

format: exec_brief

scoring: enabled

citations: required

counter_ev: on

quality_gate: 7.0

template: saving

reusable: true

product: ready ✓

live your product

Enter query...

Research

Where We Actually Sit

Three layers. Only one is easy to copy.

Cool, but Not Special

Table stakes

Multi-agent systems

Emerging pattern, widely known

Multiple AI models

Model ensembles are active research

Web search + synthesis

Every deep research tool does this

Structured outputs

JSON schemas and validation exist

Pretty Interesting, Patent Pending, Not a Forever Moat

Architecture & Methodology

Adversarial pipeline

Competing models cross-examine at every stage

Enforced counter-evidence

Every assertion must argue against itself

Confidence auto-capping

System won't let itself be overconfident

Infinite stacking

One engine, any product, no ceiling

Why this is the moat

Each decision tested across hundreds of reports. Multiply across 19 modules in Entity Intelligence — hundreds of calibrated decisions per report, all interacting.

Reports take hours to days to run. Reading the output, understanding what's good, what's bad, and what to change — that's a skill built through reps, not reading documentation.

Prompt engineering at this scale is not "write a good prompt." It's tuning how machines form views, structure arguments, weigh evidence, and calibrate confidence.

The difference between a useful AI product and a toy is in formatting, tone, section ordering, evidence hierarchy, and a hundred other things nobody notices until they're wrong. We've made those mistakes already.

Hardest to Replicate

Judgment, Execution & Testing Nuance — Our moat lives here

1 Query intent classification

2 Search string generation

3 Source screening & cutoffs

4 Relevance vs. authority ranking

5 Token depth per source

6 Depth vs. breadth tradeoff

7 Persona & voice calibration

8 Assertion strength tuning

9 Evidence hierarchy rules

10 Fact deduplication prompts

11 Attack persona aggressiveness

12 Counter-evidence search scope

Cross-Module Synthesis

13 Cross-agent fact layer — surface shared data points across modules

14 Cross-module deduplication — same fact surfaced by multiple agents

15 Section rewriting — eliminate repetitive reflection patterns

16 Narrative pattern cleanup — each module writes at full power, synthesis reconciles

17 Contradiction resolution — when modules disagree, surface the tension

18 Section ordering & flow

19 Formatting & tone polish

... per module, per product

Products Created with the Cortix Engine

Company Analysis

Entity Intelligence

Modules

Runtime

2+ hrs

Industry & Sector

Market Intelligence

Modules

Runtime

1+ hr

Deal Analysis

Tension Report

Modules

Runtime

15 min

Portfolio Analytics

Cortix Lens

Modules

Runtime

40 min

Commercial Due Diligence

Cortix Diligence

Modules

Runtime

4 hrs

Buyer / M&A Targeting

M&A Targeting

Modules

Runtime

1+ hr

AI Impact Assessment

AI Exposure

Modules

Runtime

45 min

Research Validation

Paper Validation

Modules

Runtime

1 hr

Event & Topic Research

Event Intelligence

Modules

Runtime

30+ min

Build Your Own

Cortix Builder

Modules

Runtime

∞

Anatomy of One Cortix Product

Input

Query + Docs

18 Modules

Financial

Revenue

Capital

Strategy

Operator

Dealmaker

Investigator

Legal

Governance

Reputation

Tech

Talent

Adversary

Historian

Investor

Context

AI Exposure

Compliance

↓ ↓ ↓

Synthesizer

Cross-module analysis · contradictions · exec summary

Report

Scored · Cited

Modules

LLM Providers

Search Engines

460+

LLM Calls

1,600+

Searches

Review Steps

Hours

Example: Entity Intelligence product — full company deep-dive

Go to Market

Stage 1

Beta Programs

Work side by side with analytical teams on a live project. See the power of the engine in their workflow, on their data, for their use case. Every beta sharpens modules, templates, and scoring — the product gets better with each engagement.

Target Verticals

Commercial due diligence

New deal review

Wealth Mgmt

Portfolio analytics

Consulting

Market landscape

Investment Banking

M&A targeting

How It Works

Embedded engagement

Cortix runs alongside the team's existing research process on a real deal or project

Compare outputs

Side-by-side with their analysts — what did Cortix find that they missed, and vice versa?

Refine the engine

Every beta sharpens modules, templates, and scoring — the product gets better with each engagement

AI-Native Consulting

Take engagements. High-touch advisory powered by the Cortix engine. Each engagement = revenue + case study + engine refinement. The consulting model, rebuilt on AI infrastructure.

Economics

Traditional advisory: $50K–$500K+, weeks of work. Cortix: comparable depth in minutes to hours. Marginal cost = API + search.

Traditional

$50K–500K

Cortix

Minutes

Product Company

Sell subscriptions. Self-serve products that pay you per report, per seat, per use case. Lower friction, broader reach. Scale without headcount.

Pursuing both businesses. Beta programs will refine the model for each.

Open Question: Studio

Can we get inside research houses that want to build targeted research products? Third parties building on the Cortix engine — their templates, their domains, our infrastructure. Platform economics at scale.