AI Architecture · 30+ AI Features

AI that shows its work.

Name: Versatum
Availability: OnlineOnly
Author: Versatum

No black boxes. 30+ AI features — including an autonomous Trader fleet and a tool-calling Copilot Agent — analyze, debate, and explain the recommendations they make. Two production models do all of it, and every page tells you which parts of it are rules rather than a model.

Try the AI See All Features

LLM Models in Production

Deep Research Roles

ML Factors

Asset Classes

Two models, not a menagerie: one text model behind every written analysis and one vision model for reading chart images. Our provider routing pins an fp8 quantization floor, so a cheaper 4-bit copy of the model is never served in its place — of the 30 endpoints offered for our default model, 11 clear that floor. Every model reply is parsed through a repair layer before it reaches you, and a value we could not measure is labelled unmeasurable rather than printed as a zero.

AI Copilot

Ask your data anything.

Every widget, chart, and data point on the platform is queryable. The Copilot runs on the same text model as the rest of the platform — what makes it specific to you is the context we assemble for it: your current page, your watchlist, your trading style, and your portfolio.

Four focus tiers — widget, then tab, then page, then intelligent routing
Adapts responses to your trading style (day trade, swing, position)
16 widget sources and 42 page contexts wired in, across 84 dashboard widgets
Ask about any stock, option, forex pair, crypto asset, or signal in plain English

The meter is stated, not hidden. Every tier gets the same 32,768-token output budget, and the free tier is capped at 20,000 tokens a day, 10 Copilot runs and 1 Chart Patterns run — counted atomically and reset at UTC midnight. Paid tiers lift the daily token cap.

Versatum CopilotAI

What's the bull case for NVDA right now?

From the last Deep Research run, 4 of the 5 analysts are constructive. Market Agent has a breakout above resistance at $142, Fundamentals flags margin compression, News Agent reads the supply headlines as already priced. The Bull/Bear debate resolved bullish with medium confidence — here is each step...

Show me options strategies for my swing trade style

For your swing profile (2-10 day hold): bull call spread at $140/$150 (Mar 21 exp), risk-reward 1:3.2. Trade Validator scores it 71/100, which lands in the BUY band, and IV percentile is 42% — favorable for debit spreads...

Agentic AI

From analysis to action.

Two agentic systems take Versatum beyond explanation — autonomous traders that run continuous strategies, and a Copilot that plans and acts. Both are Pro+ features, and both are built to be left alone: eight of our long runs survive navigation, so a 30-to-90-second scan keeps working if you change page mid-run.

Autonomous Trader

Autonomous Trading Fleet

Configure agents — US Daily, US Hourly, and Crypto — with capital and risk, then run them in backtest or continuous live mode. Progress reports day-by-day with phase badges and a FIFO-matched trade ledger.

3 agent types
Per-agent risk tuning (1–10)
$10,000 default starting capital
Mirrors fills to your paper account

Copilot Agent Mode

Tool-Calling Deep Agent

Toggle Copilot from chat into a deep agent. It plans multi-step tasks, executes shell commands in a sandbox, generates downloadable files, and uses vision on attached charts — showing every plan card and tool call as it works.

Multi-step plans visible as todo cards
Sandboxed shell + file tools
166 MCP tools available to the agent
Persistent threaded sessions

Deep Research

13-role research pipeline.

Five specialized analysts run in parallel, then a two-round Bull/Bear debate, a Research Manager, a Trader, and a 3-debater Risk team deliver a final BUY/SELL/HOLD with confidence. Thirteen distinct roles, six stages, eighteen model calls — every reasoning step visible.

Market Agent

Price action, technicals, market structure, support/resistance levels

Sentiment Agent

Social media pulse, options flow sentiment, retail vs institutional positioning

News Agent

Breaking news, narrative shifts, media momentum, event catalysts

Fundamentals Agent

Financials, valuations, growth metrics, competitive moat analysis

Earnings Agent

Transcript NLP, guidance changes, estimate revisions, surprise history

Risk Team — 3 debaters

Two rounds over downside scenarios, correlations and tail risk, then a Risk Manager signs the decision

Synthesis Engine

All thirteen role outputs are synthesized into one report with confidence, risk assessment, actionable trade ideas, and full reasoning chains. Each stage has its own versioned prompt from a shared registry of 35, so you can see what every role was actually asked. Crypto runs the same shape with 12 roles and 4 analysts; forex runs a shorter 3-stage version, also with 4 analysts.

Adversarial Debate

Bull vs Bear — the case gets challenged.

Inside Deep Research, the analysts' work goes through two rounds of adversarial debate before anything is recommended. A Bull analyst builds the case, a Bear analyst attacks it, and a Research Manager rules on it. The Risk team then debates the plan for two more rounds before the decision is signed.

Bull Analyst

Builds the strongest case for the trade — catalysts, upside targets, momentum signals

Bear Analyst

Stress-tests every weakness — overvaluation risks, competitive threats, macro headwinds

Research Manager

Reads both rounds, picks a side with a confidence, and hands it to a Trader who turns the call into a plan

Adaptive AI

AI that matches your style.

Select your trading style and every AI feature recalibrates — signals, thresholds, timeframes, risk parameters, options strategies, and the prompts themselves all adapt automatically. Works across all four asset classes: stocks, options, forex, and crypto.

Multi-day holds. Technical setups are balanced against fundamental catalysts for 2-10 day positions.

4-hour bars, 90-day history

Technical + fundamental blend

Moderate risk tolerance

Pattern & catalyst signals

Every AI capability, the Copilot, and all signal/research pipelines across stocks, options, forex, and crypto dynamically adjust when you switch trading styles.

Complete Suite

30+ AI features, one platform.

From an autonomous Trader fleet to chart vision, from RL agents to earnings NLP — every output shows the inputs behind it, and the ones that are pure rules say so.

Agentic

Autonomous Trader

Three agent types — US Daily, US Hourly, Crypto. Give one capital and a risk setting, run it in backtest or continuous mode, and mirror its fills to paper.

Agentic

Copilot Agent Mode

Tool-calling deep agent with multi-step plans, sandboxed shell execution, file generation, image input, and 166 MCP tools.

Multi-Agent

Deep Research

13 distinct roles across 6 stages and 18 model calls — 5 parallel analysts, a two-round Bull/Bear debate, Research Manager, Trader, a 3-debater Risk team, and a final Risk Manager.

Scoring

Trade Validator

16 stages in 5 waves score 12 weighted dimensions, then 14 gating rules and one synthesis call land a verdict in five bands with risk-adjusted sizing.

Factor Model

Stock Picker

Percentile-ranks every name against the universe on 35 factors in 7 categories — value, quality, growth, financial health, momentum, earnings quality, analyst sentiment — then blends them by weight.

RL Training Lab

PPO, Rainbow DQN, and IQN reinforcement learning agents trained on market data to discover optimal strategies.

Deterministic

Investment Committee

Eight investor personas — Buffett, Burry, Munger, Wood, Graham, Ackman, Lynch, Fisher — plus four quant analysts score the same name by rule. Zero model calls.

Pipeline

Morning Picks

Before the bell, 300 large caps are scored, the top 70 fan out to validation, and up to 20 setups publish with verdict, entry, stop, targets, and size. Five slots roll through the day.

Scanner

Gap Analyzer

247 large caps polled for moves of 0.75% or more; the twelve biggest go through one batched call that calls each PANIC, JUSTIFIED or UNCLEAR and cites the headlines behind it.

Options

Options Signals

AI-scored put/call signals with Greeks, IV percentile, max pain, GEX, unusual flow, and a return simulator.

Options

Options Screener

Up to 10 tickers per run, with portfolio Greeks, cross-ticker strategies, and risk aggregation.

Vision

Chart Patterns

A vision pass reads the rendered chart for 14 classical patterns with confidence, completion %, target, and S/R levels. Anything under 50/100 confidence is discarded.

NLP

News Sentiment

News polled every 15 minutes, with multi-dimensional sentiment scoring, narrative arc tracking, and theme clustering. An untagged article stays untagged — never counted as neutral.

Scanner

Factor Alpha

Cross-stock alpha ranking across momentum, volume anomaly, volatility setup, mean reversion, and trend strength.

Sentiment

Fear & Greed

Per-stock sentiment composite with breadth, news, volatility, volume, and AI contrarian analysis.

Flow

Whale Tracker

Volume microstructure for equities — A/D score, OBV, institutional vs retail split, and divergence alerts.

Analytics

Watchlist Health

Per-symbol 0-100 health scores with a hold / add / trim / remove verdict and the issues behind each one.

Risk

Risk Mapper

Your watchlist as a correlation map — pairwise matrix, diversification score, sector mix and average beta vs SPY across the names you follow, each with its measurement coverage in names.

Prediction

Economic Calendar

AI impact predictions for FOMC, CPI, NFP, earnings seasons, and geopolitical catalysts.

Analysis

Compare Stocks

Side-by-side multi-dimensional comparison across fundamentals, technicals, sentiment, and valuation.

Forex

Session-aware forex with pip-based signals, a 4-session clock, and a 4-analyst research pipeline that runs in 3 stages.

Crypto

11-tab crypto suite with Fear & Greed, Whale Detector, Alpha Scanner, Narrative Radar, and 4-analyst deep research.

Trade Scoring

Confluence scoring.

Trade Validator runs 16 stages in 5 waves and blends 12 weighted dimensions into one score, then applies 14 gating rules and lands it in one of five verdict bands. Exactly one of those stages calls a model — the head-of-desk synthesis at the end. Everything feeding it is arithmetic you can check.

Weights are per trading style: day traders get momentum, VWAP and intraday flow; position traders get fundamentals and analyst targets. A dimension we could not measure for a ticker is dropped from the denominator rather than scored zero, so a thinly covered name is not quietly punished for our missing data.

Swing-trade weights · sums to 100

Chart Pattern

Momentum & Technicals

Sector / Industry Rotation

News Sentiment

Analyst Price Target

Market Regime

Fundamentals & Valuation

Gap Analysis

Eight investors, and no model at all.

Each persona is a set of published criteria turned into arithmetic — value, growth, contrarian, activist, GARP, quality — scoring the same name independently. Alongside them, four quant analysts (fundamentals, valuation, sentiment, technicals) score it too. Twelve independent signals, and a portfolio manager rolls them into an action, a confidence and a position size.

Warren Buffett

Value investing, economic moats, long-term compounding

Michael Burry

Contrarian deep value, macro shorts, asymmetric bets

Charlie Munger

Quality at fair price, mental models, patience

Cathie Wood

Disruptive innovation, growth at scale, thematic ARK bets

Benjamin Graham

Margin of safety, intrinsic value, deep-value screens

Bill Ackman

Activist conviction bets, catalyst-driven, concentrated positions

Peter Lynch

GARP, invest in what you know, tenbagger hunting

Phil Fisher

Scuttlebutt research, growth quality, long compounders

On an explainability page, this part has to be said plainly

The committee makes zero model calls. The personas do not debate, they do not converse, and there is no language model anywhere in the run — it is deterministic scoring in the style of eight famous investors, which is why the same inputs always produce the same committee. Do not count these eight toward any tally of AI models; the only model-derived input that reaches the committee at all is cached news sentiment, and an article we never managed to tag is dropped rather than counted as neutral.

Machine Learning

Reinforcement learning agents.

Three RL architectures learn trading policies from simulated episodes on historical data. Train, backtest, and deploy — all within the platform. No language model is involved: these learn a policy from rewards, they do not write prose.

PPO

Proximal Policy Optimization

Stable, reliable policy gradient method for continuous trading decisions. Balances exploration with exploitation.

Rainbow DQN

Rainbow Deep Q-Network

Combines 6 DQN advances — prioritized replay, dueling architecture, noisy nets, and distributional RL.

IQN

Implicit Quantile Network

Models the full return distribution for risk-aware decisions. Understands tail risk, not just expected returns.

Plus: the 35-factor Stock Picker

Not a neural net and not a black box: each of 35 factors across 7 categories — value, quality, growth, financial health, momentum, earnings quality, analyst sentiment — is percentile-ranked against the rest of the universe, then blended by weight. A custom strategy builder lets you re-weight them to match your thesis, and how many are active depends on your trading style: 10 for day trade, 25 for swing, 18 for long term. Each score carries a confidence equal to the share of factor weight that actually had data behind it.

Vision AI

Chart pattern detection.

The vision model is handed a rendered chart image and asked for 14 classical patterns, each with a confidence score, a completion percentage, a target and the support and resistance it read. A run is three model calls — two vision passes and one text pass — and any pattern below 50 out of 100 confidence is thrown away rather than shown to you.

When a chart read fails, we say so. Morning Picks records the reason, shows a "Some analyses failed" banner with the count, and the synthesis prompt is explicitly told not to infer a chart it never saw — so a briefing built without chart input can no longer be mistaken for a healthy one.

Head & Shoulders

Double Top/Bottom

Triangles

Channels & Flags

Cup & Handle

Wedges

Chart Patterns

AAPL|AI Vision

Chart Render

Candlestick Chart

AAPL3M

Patterns Found

Head & Shoulders

bearish92% complete

87%

Rising Wedge

bearish78% complete

71%

Bull Engulfing

bullish100% complete

64%

Key Levels

Support

$188.5

$185.2

$180.0

Resistance

$196.8

$201.5

$205.0

Trade Signal

SELL

Confidence82%

Entry$192.40

Target$185.20

Stop$197.50

R:R1:1.4

H&S neckline break confirmed

Volume declining on right shoulder

Target = head-to-neckline distance

Vision AI|14 pattern types

1/4

Chart Signals · No AI

The one page where we chose not to use AI.

Three studies, written line by line from their reference source and run on our own bars. No model touches Chart Signals — every number is deterministic, reproducible and pinned to a committed source hash. Six bar intervals, seven lookbacks, and it works on stocks, crypto and forex.

Follow-Through Levels

Up to ten historical-frequency price levels — five above the prior high, five below the prior low — each with its own probability reading, fill bands, and a WIN / LOSS / profitability panel. Levels that measure 0.00% are hidden, so the count is whatever the data supports rather than a fixed number.

Money-Flow Trail

A per-bar trailing stop-and-reverse line, its two-bar-lag twin, and a band between them that flips colour with the trend. BUY and SELL tags appear only once a flip is confirmed.

Indicator Consensus

Named after the published script it came from, and it is not one: it is a weighted scoreboard. 29 switchable contributions over 16 computed indicators — 10 of them paying at defaults — plus a five-EMA ribbon, a Hull average, and a 52-cell panel naming every contribution currently paying.

Zero model calls

Nothing here is sent to a model. The same bars always produce the same reading, which is not something an LLM can promise you.

Pinned to its source

Each study prints its script revision and a 12-hex hash of the committed reference file it was written from, next to the reading itself.

1,637 contract checks

Plus a 379-mutation sweep that deliberately breaks the maths to prove those checks would notice. All 26 places where we knowingly deviate from the original scripts are listed on screen.

The view cannot move the number

All 55 inputs are classed compute, window or presentation, and every control tells you which — so changing what you are looking at provably cannot change what it reads.

You will find it under Validate, beside Trade Validator and Chart Patterns — the answer to "is the chart confirming my idea?" that you can audit yourself, with no tier gate and no daily cap.

Built so it cannot lie to you.

Try the AI See Pricing