Averum
Volume I · Cover
--:--:-- Investor Brief · 2026
Averum · Investor Brief · 2026
Averum
The governed agent runtime
for finance.
AI agents are entering markets. Money needs a runtime.
Live in production First-party strategies running Broker-bound · IBKR Audit · replay · introspect
Thesis · 02

Generic AI runtimes are landing.
None of them were built for markets.

Problem · 03

Everyone is building AI advice.
Almost nobody is building AI control.

A capable model with a broker key is one bad afternoon away from a blow-up. Not because the model is dumb — because nothing around it is built for what happens when it's wrong.

Incident · the duplicate-fire
The market data WebSocket stutters for 800 ms. The agent retries the same "BUY 100 SPY" three times because the position read came back stale.
What has to stop it
  • idempotent OMS with proposal IDs
  • position state agents can trust mid-flight
  • fill-event triggers route to a different flow than entry
Incident · the size mistake
Agent reads a 5-contract suggestion as 500 because today's volatility prompt was three times longer than yesterday's. Stop logic kicks in. The damage is already done.
What has to stop it
  • capital tier & max-position caps as data, not prose
  • pre-trade risk gates the agent can't override
  • buying-power check before order proposal, not after
Incident · the LLM that died
Mid-trade, the model endpoint 504s. The agent has an open position, a working stop, and no way to think. What runs?
What has to stop it
  • deterministic exit rules independent of the LLM
  • halt + manual-exit triggered by health checks
  • capability downgrade · supervisor takes the wheel
Incident · the silent drift
The strategy works for six weeks, then quietly stops. Win-rate decays from 0.58 to 0.41. Nobody can say which prompt edit, which model swap, or which regime change started it.
What has to stop it
  • content-hashed decision trace · linked to outcome
  • replay against original market state
  • adversarial judges before any change ships

Each line above is a runtime primitive on the next slide. They exist because each one is a specific thing that has gone wrong in production.

What markets demand · 04

Six assumptions general agents make.
All of them break in markets.

The runtime absorbs what the model can't — so the model can do what it's good at.

01
Assumption
The world is static between observation and action.
Markets demand
Freshness as a first-class input.
02
Assumption
One cadence — "someone pinged me."
Markets demand
A trigger plane across time scales.
03
Assumption
Pure input → output. The agent decides.
Markets demand
A stateful world the agent can break — and the runtime can't.
04
Assumption
All actions reversible. Re-roll the prompt.
Markets demand
Irreversibility-aware approval.
05
Assumption
Logging the API call is the audit.
Markets demand
A causal chain — replayable against the original market.
06
Assumption
One agent. One voice. One model.
Markets demand
Composition with conflict resolution.

Plus one more: 70% bullish and 95% bullish are different positions — not different phrasings.

Primitives · 05

Finance needs something different.

Six primitives · zero shipped by horizontal runtimes
In · the world
What the agent sees.
01
Demand-driven market data
Agents declare what wakes them — not every tick. Watchplans, freshness policies, economic calendar.
02
Trigger routing
Signal · OMS event · chat · schedule — different wakes route to different flows. Don't run the entry agent on a fill.
03
Market narratives
Market state packaged for LLM consumption — the picture an agent needs, without flooding the context window.
Averum runtime ▸
The finance-native runtime for agents that operate around real money.
Out · the action
What it can survive.
04
Domain tooling
Order proposal, position close, OMS introspection, runtime state, flow composition — permissioned per agent, audited per call.
05
Tested connectivity
When the LLM endpoint dies with an open position — capability flags, halt + manual-exit, deterministic safety rails.
06
Replay that doesn't bankrupt you
Intelligent freezing — mirror prior agent outputs, re-invoke only what's changed. Reconstruct any decision against original market state.
01–03 what the runtime feeds the agent  ·  04–06 what the runtime catches when the agent is wrong
Control modes · 07

Manual. Assist. Autonomous.

The same runtime serves cautious users, active traders, professional desks, and embedded broker experiences.

Mode 01
Manual
The user drives. Agents watch, explain, and answer. Every order is yours.
Mode 02
Assist
Agents propose actions. The user approves, rejects, resizes, or asks for more reasoning.
Mode 03 · Default for desks
Autonomous
Agents act inside explicit limits, capital tiers, policies, and broker permissions.
Always-on controls
  • Risk, size, symbols, and strategy bounds
  • Human approval when rules require it
  • Position and exposure awareness before action
  • Full history of why every decision happened

Averum lets autonomy expand only as trust, policy, and evidence allow.

Regulation · 08

When the rules arrive, Averum is what they describe.

Brokers become the forcing function before regulators do.
We're already operating to the standard institutional algo trading has lived under for fifteen years.

Today
Today's gap
No specific framework governs AI agents executing trades in retail accounts. The existing regime covers institutional order flow — assuming human-in-the-loop or pre-trade risk gates, not autonomous agent reasoning.
SEC 15c3-5 FINRA 3110 / 3120 MiFID II Art. 17 EU AI Act · high-risk
Next 24–36 months
What brokers will demand · what regulators codify
  • Capability constraints per agent · per account · per capital tier
  • Mandatory audit, replay, and supervisory frameworks
  • Kill switches + human-approval thresholds
  • Disclosure of AI involvement in trading decisions
  • Written supervision policies for autonomous workflows
Already shipped
Averum is operating to the standard before it lands
shipped
Capability flags
Per agent · per account · per capital tier.
shipped
Immutable audit ledger
Content-hashed · tamper-evident.
shipped
Pre-trade risk gates
Approval policy at the runtime.
shipped
Replay against market state
Reconstructable, reviewable.
shipped
Structured introspection
Natural-language "why did we trade?"

We are what compliance will look like — before regulators write the rule.

Audit + introspection · 09

Most trading systems can log what happened.
Averum is built to explain why — and improve the system.

01 · AUDITReconstruct
02 · ASKInterrogate
03 · STRESSAdversarial review
04 · IMPROVERewrite the agent
01
Reconstruct every decision.
Market context, prompts, tools, agent outputs, approvals, orders, fills, outcomes — content-hashed and immutable.
02
Ask in natural language.
"Why did we enter?" · "What evidence mattered?" · "Which agent disagreed?"
03
Adversarial review.
"Will this overfit?" · "Is this change really necessary?" · "Does it survive a regime shift?"
LLM judges challenge proposed changes before they ship.
04
The agent rewrites itself.
Surviving changes ship — prompts, tools, gates, thresholds, policies — against explicit targets.
Every loop, the agent gets sharper.

Decisions become explainable. Explainable decisions become testable. Tested decisions become better strategies — automatically.

Why now · 10

The window is opening now.

Capability, demand, and governance pressure are all bending up at the same time. We are arriving the year these lines cross — and the urgent vertical has no incumbent.

Adoption · capability · pressure
202020222024we are here · 202620282030
2020
LLMs reach knowledge-worker quality
2023
Tool use, function calling, multi-agent demos
2025
Inference fast + cheap enough for intraday agentic decisions
2026 · now
Horizontal agent runtimes move from demos to deployment.
Brokers + funds need a finance-native answer.
2027–28
Brokers demand audit/replay/capability gates · regulators codify
2030
Tokenized · 24/7 · agent-mediated markets are the default

The horizontal AI runtime wave creates the category. Finance creates the urgent vertical.

Market + wedge · 11
$10B+ agent-runtime TAM in finance by 2030.
Roughly 15–20% of the $50B+ horizontal agent-runtime market — financial services is consistently the largest enterprise software vertical by spend.
Bottom-up addressable
Segment Addressable buyers ACV Segment TAM
Professional desks · RIAs · prop shops · family offices ~30,000 in the West $50K avg $1.5B
Brokers · fintechs · platforms (embed) ~250 globally $1M+ $250M+
Active retail / prosumer traders ~25M globally ~$80/mo blended ~$24B/yr
Wedge — start with operators who already supervise money
Prop desks · family offices · RIAs · active trading groups · broker-adjacent fintechs.

They need agent automation, but cannot run black boxes. High urgency. High willingness to pay. Low tolerance for missing audit.

Annual contract band
$15K–$300K+/ year
01
Already run trading workflows.
No category education needed — agents replace or augment tools they use today.
02
Already need policy, supervision, records.
Governance is a feature, not a tax — it maps to obligations they have today.
03
Can pay before consumer scale.
$15K–$300K+ annual contracts. Seat economics work without prosumer.
04
Create the proof brokers need.
Reference deployments become the wedge into embedded distribution.

100 customers at $50K ARR is a $5M business. 1–2% blended capture lands $100M+ ARR. Broker embedding is the path.

Go to market · 12

Professional proof → platform embed → prosumer scale.

Now
STEP 01 · PROFESSIONAL DESIGN PARTNERS
5–10 prop desks, family offices, active trading groups, RIAs, broker-adjacent fintechs.
Real capital workflows. Replayable decisions. Compliance feedback. Reference customers.
GoalLand the first paying ICPs and prove supervised autonomy.
6–12 months
STEP 02 · BROKER + PLATFORM EMBED
Turn professional proof into distribution.
Averum becomes the governed agent layer brokers, fintechs, charting platforms, and horizontal runtimes can offer without building it themselves.
GoalEmbedded runtime partnerships. White-label or co-branded Navigator.
After platform proof
STEP 03 · PROSUMER SCALE
Navigator launches as the retail-facing surface — once governance, workflows, and trust are hardened.
Active trader adoption. Subscription revenue. Usage-based AI economics. Consumer-grade command center on enterprise-grade infrastructure.
GoalScale through trusted distribution, not retail hype.

Not "Robinhood with an agent." The governed agent runtime brokers, desks, and serious traders can trust.

Competitive landscape · 13

Incumbents own pieces.
Nobody owns the chain.

Category Examples Why they don't ship Averum
Horizontal agent runtimes Cloud agent platforms · open-source agent frameworks · model-vendor SDKs Generic by design. No market-native OMS, broker layer, or finance-shaped governance.
AI platforms Foundation-model providers Not market-native. No execution surface or broker-bound audit loop.
Brokerages Retail brokers · introducing brokers Distribution and accounts, but legacy stack and conservative agent-risk posture.
Trading systems Institutional terminals · OMS / EMS vendors Strong workflow infrastructure, but not agent-native or prosumer-ready.
Quant tools Backtesting platforms · screeners · retail quant tools Good at research, weak on governed agent execution and natural-language introspection.
Consumer AI trading apps Chatbots · signal apps · copy-trading tools Advice-first. Shallow audit, often not broker-native.

Averum owns the chain — market context → agent decision → broker action → replayable outcome.

Partner or build · 14

Brokers and funds have reasons to partner.

Brokers
They have accounts, trust, execution rails, and distribution.
But they also have:
  • Legacy stacks · conservative compliance posture
  • Slow product cycles
  • High downside if an internally built agent layer fails
  • Limited appetite to own model, runtime, and strategy risk from scratch

Averum gives brokers a governed AI layer they can embed — without betting the company on an internal moonshot.

Funds + professional operators
They may build proprietary agent workflows.
But they're not usually building:
  • Reusable runtime primitives
  • Broker embed surfaces
  • Prosumer-grade UX
  • Cross-strategy audit and introspection tooling
  • Continual evaluation and self-improvement infrastructure

Averum becomes shared infrastructure around many strategies — not one firm's internal tool.

Business model · 15

Four tiers. Each earns its price.

Tier Price Customer What they get
Navigator $29–79 / mo Retail AI trading command center with audit and approval controls.
Navigator Pro $199–699 / mo Active traders Multi-agent, multi-account, custom rules, deeper replay.
Desk $15K–75K / yr Prop desks · family offices · RIAs Embedded runtime, custom strategies, governance, support.
Enterprise $75K–300K+ / yr Funds · brokers · fintechs White-label, broker integrations, compliance support, deployment options.
Comparable benchmarks · single-seat or per-account
TradingView
$15–60 / mo
anchor for Navigator retail tier
Trade Ideas · TrendSpider
$48–228 / mo
anchor for Navigator Pro
FactSet · Bloomberg · Refinitiv
$12K–30K / yr
anchor for Desk + Enterprise tiers

Expansion drivers: subscriptions · enterprise licensing · broker distribution · usage-based AI economics · embedded partnerships.

Financial model · 16

Path to $100M+ ARR — built customer by customer.

Illustrative model. Each year is the same arithmetic — count of customers in each tier, tier ACV, and broker-embed deals — not a top-down hope.

Year 1
Land the wedge
Desk pilots8×$45K$0.36M
Enterprise pilot1×$150K$0.15M
Navigator beta1.5K×$49/mo$0.9M
Pro300×$299/mo$1.1M
Base ARR $2.5Mrange $2–4M
Upside · 1 broker LOI lands → $5M+
Year 3
Scale + first embed
Desks220×$55K$12M
Enterprise22×$180K$4M
Broker embed2×$1.5M$3M
Navigator15K×$59/mo$10.6M
Pro2.5K×$349/mo$10.5M
Base ARR $30Mrange $25–35M
Upside · 4–5 broker embeds → $50M+
Year 5
Plane of distribution
Desks1,500×$62K$93M
Enterprise75×$220K$16M
Broker embed10×$2M$20M
Navigator90K×$59/mo$63M
Pro12K×$399/mo$57M
Base ARR $180Mcapture math · $100M+ floor
Upside · 20+ broker embeds, prosumer scales → $300M+
Where the dollars come from

~70% of Year 5 base ARR comes from professional + enterprise + broker embed — the wedge and its distribution. Prosumer is the lever, not the foundation.

Margin shape

Gross margin 70% Year 1 → 75%+ thereafter. LLM inference declines as a % of revenue as we mature prompt + cache strategies and ladder usage-based pricing on top.

Math vs. capture

Year 5 plan equals ~5% of professional, 4% of broker / platform, 0.4% of prosumer against the bottom-up TAM on slide 12. We are sized to be a credible category leader, not a corner of a market.

Moat · 17

The moat isn't one model. It's three layers stacking.

One is structurally true today. One compounds with every governed action. One we are earning, openly — and the regulatory wedge is what unlocks it.

Built · today
01
Vertical depth — we own the chain.
Horizontal agent runtimes structurally cannot reach orders, fills, positions, or replayable market state. They sit one layer above the broker; we sit between the agent and the broker. That gap doesn't close with more inference — it closes with seven primitives nobody else ships together.
Market datademand-driven · 50 Hz · narratives
Agent runtimetools · memory · structured outputs · hooks
OMS + brokerproposed orders · fills · capability flags
Risk + auditgates · content-hashed ledger · replay
live
Compounding · year 1+
02
Decision corpus — we know what works.
Every governed run produces a content-hashed trace — context, prompts, tool calls, agent disagreement, approvals, orders, fills, outcomes. It accrues with usage, links cause to consequence, and feeds straight into self-improvement: prompts, gates, thresholds, and policies are tuned against ground-truth outcomes nobody else can see.
Trace shapedecision → action → outcome, content-hashed
Used byreplay · adversarial judges · target-driven evals
Asymmetrygrows with every customer · not synthesizable
Resultwe don't just improve your strategies — we know what works and what doesn't
compounding
Earning · year 2–5
03
Distribution becomes gravity — where the rules and partners land.
We don't have signed partners yet. We are building toward this layer openly. The regulatory wedge (slide 9) is what turns it into a moat: brokers must underwrite a control layer before opening their surface to AI agents. The first runtime to clear that bar at one broker becomes the default for the rest — the underwriting work doesn't repeat.
Forcing functionbrokers' downside risk · then regulator codification
First wedgedesign partners · supervised desks (slide 14)
Switching costcapability gates · approval policy · audit history
Statusnot yet earned · this round funds it
earning

Horizontal runtimes don't sit between the agent and the broker. They don't see the data that matters — and they don't pass the underwriting brokers will require.

Team · 18

Institutional depth. Consumer vision. Operational backbone.

Co-founder · Runtime + domain
Adam Wielowieyski
Goldman Sachs (credit derivatives) · JPMorgan (pricing systems) · Head of Analytics + Data Strategy, Hong Kong Stock Exchange · Imperial College London (CS) · HKUST (MBA).
Built institutional trading and data infrastructure — knows what "governed" actually means in production.
Co-founder · Product + GTM
Seth Alsbury
6 startups · $300M+ raised · EIR at Target Corp / Roundel · UC Berkeley (Honors).
Consumer product instinct + serial fundraising experience.
Co-founder · Fintech + ops
Steve Rubenstein
Serial fintech founder · sold first company at 25 · multiple retail and commercial banking ventures · fintech advisory and regulated-market operating experience.
Operational rigor and regulated-fintech reps.
Advisor
Quinton Alsbury
Amazon · Saxon. Product architecture, enterprise AI, distribution.
The ask · 19

Seed round to land the wedge.

40%
EngineeringNavigator GA · broker integrations · audit/introspection · self-improvement harness · governance hardening.
30%
Design partners + GTMFirst 10 paying customers · professional operator pilots · broker and platform conversations.
20%
Compliance + regulatoryCounsel · BD/RIA structure · audit infrastructure · supervision framework.
10%
ReserveHeadroom for opportunistic hires and partnerships.
Asking $[FILL: X]M  ·  Lead $[FILL: Y] committed / in discussion.

This round buys us to Year 1 ARR and the first broker / platform proof.

Bloomberg was built for humans in markets.
Agent runtimes are built for agents in software.
Averum is built for agents in markets.
Robinhood made trading simple. Averum makes AI trading controllable.
Live now · navigator running · cycle 4128
Backup · A

Ask the system why.

Useful for demos and partner conversations. The introspection surface is real — these are queries you can run today against the audit ledger.

Why & what
  • Why did Omega approve this entry?
  • What market context changed between proposal and fill?
  • Which expert disagreed and why?
  • Did the decision violate any current rule?
Counterfactual & adversarial
  • Would the strategy make the same decision under today's prompt?
  • Which feature had the highest influence?
  • What would have prevented the losing trade?
  • What changes would an adversarial reviewer recommend?
Backup · B

From decision to better strategy.

The loop
  1. 01 Capture the full decision trace.
  2. 02 Link the decision to its market outcome.
  3. 03 Rerun under controlled experiments.
  4. 04 Apply adversarial LLM judgement.
  5. 05 Compare against explicit targets.
  6. 06 Propose updates — prompts, inputs, flows, thresholds, policies.
  7. 07 Promote only changes that pass evaluation.
Targets
PnL Drawdown Sharpe Win rate False positives Missed trades Execution quality Rule adherence

Self-improvement as a disciplined harness — not vague "the AI learns."

Backup · C

Why partner instead of build.

Build internally
  • Own agent-risk from scratch
  • Recruit AI runtime, trading systems, broker integration, compliance, UX expertise
  • Build audit, replay, evaluation, introspection
  • Maintain model/provider changes
  • Wait through long internal product cycles
Partner with Averum
  • Embed a governed agent runtime
  • Keep broker / customer relationship
  • Control policy and approval model
  • Launch faster with a finance-native stack
  • Benefit from ongoing runtime + evaluation improvements
Backup · D

Built, not imagined.

Use only when investors ask how much exists today. Replace placeholders with current evidence before sharing.

Live or recent screenshots
Navigator / Council surface · timestamp · IBKR connection state.
Live decision count
Recent decisions · cycle history · per-strategy metrics.
Market data demand graph
Watchplan evidence — what's subscribed, why, since when.
Flow execution trace
Forward-propagating flow audit · one execution shape across DAG and tool-agent.
Observability stream
Dashboard event timeline · framework + trading + market streams.
Audit / introspection query
Once the first cut lands, "why did we trade?" answers in natural language.

A production codebase with live market data, broker connectivity, flow execution, observability, and first-party strategies. Navigator is the product surface being built on top.

Backup · E

Why we believe brokers and regulators will require this.

Today's gap

No specific framework governs AI agents executing trades in retail accounts. The closest reference points all assume human-in-the-loop or pre-trade risk gates — not autonomous agent reasoning.

SEC 15c3-5 · market access FINRA 3110 / 3120 · supervision MiFID II Art. 17 · RTS 6 EU AI Act · high-risk
What's coming
  • Capability constraints per agent and per account
  • Mandatory audit, replay, supervisory frameworks
  • Kill switches and human-approval thresholds
  • Disclosure of AI involvement in trading decisions
  • Written supervision policies for autonomous workflows
Why brokers move first

Brokers cannot let agents touch retail capital without this layer. The downside risk — an "OpenClaw" agent going wild on customer accounts — lands on the broker's balance sheet, reputation, and license. They will demand Averum-class controls before opening their surfaces.

Regulation is timing arbitrage. We are on the right side of it.

Backup · F

The agentic trading stack.

The runtime primitives required to run governed financial agents.

01
Market Data Substrate
Demand-driven watchplans, triggers, event subscriptions; tick pricing up to 50 Hz, bars, indicators, options.
02
Agentic Context
Feature recipes, market narratives, event context, context builders that turn raw market state into LLM-usable intelligence.
03
Agent Runtime
Tool use, memory, prompt assembly, model routing, structured outputs, lifecycle hooks, cost tracking — deterministic around LLM calls.
04
Flow Execution Fabric
Forward-propagating flow graphs with transform, decision, agent, and nested-flow nodes. One execution driver, one audit shape.
05
Risk + Governance
Limits, account policy, capital tiers, permissions, safety rails — applied as data, replayable.
06
OMS + Broker Layer
Proposed orders, execution, fills, positions, broker capability enforcement.
07
Audit + Introspection
Reconstruct every decision and ask why it happened — content-hashed, immutable, queryable.
08
Self-Improvement Harness
Adversarial judges, experiment reruns, target-driven strategy updates.

Not a model. Not a signal app. Not a broker. The runtime substrate and control layer that financial agents need.