VegaForge

AI systems / CSAS

AI systems that do the work.
With you in control.

We build AI agents for support, operations, financial monitoring and prediction market research. Retrieval, tools, evaluations and human approval, built into the systems your team already uses.

How it works

01 / Where it fits

A useful agent is a system, not a prompt.

The model is one part. The rest is data, permissions, tools, edge cases and the interface around them.

01

Customer support

Answer from the real knowledge base, hand off edge cases with context, and keep every answer auditable.

02

Internal operations

Read the documents, check the systems and prepare the next action. Your team approves before anything changes.

03

Research and monitoring

Watch sources, compare signals and deliver a short brief instead of another dashboard nobody opens.

04

Product AI features

Retrieval, tools, evaluations and interfaces that survive real users, not a demo that only works on the happy path.

The pipeline

Sources to action

The model sits in the middle. The rest is the system.

01

Sources

Documents, ledgers, feeds

02

Normalize

Clean, index, permissions

03

Tools

Deterministic calls

04

Agent

Reason, plan, explain

05

Approval

Human decision

06

Action

Write, send, act

02 / Two workflows

Where the work is not a chat window.

Two patterns we keep coming back to: numbers that must reconcile, and decisions that need evidence.

Financial data

Monitor, reconcile, explain.

Statements, invoices, bank feeds and market data arrive in different formats. The agent normalizes them, runs the deterministic math, reconciles what changed and writes a short morning brief. Anything unusual gets a source and an owner.

What it does

  • Normalizes ledgers, invoices and feeds.
  • Runs calculations outside the model.
  • Flags anomalies with links to the source.
  • Prepares a brief for a human to approve.
Deterministic mathSource-linked alertsAudit trailHuman approval

Prediction markets

Research the edge. Keep the risk bounded.

The agent reads the market, the resolution rules and the news, then produces a probability estimate with the evidence behind it. It compares that estimate to the price, sizes the gap against explicit limits and asks before acting.

What it does

  • Scans markets and resolution rules.
  • Builds a case with sources and counterarguments.
  • Compares the estimate with the market price.
  • Reports the tradeoff, not a certainty.
ResearchCalibrationRisk limitsHuman approval

03 / The stack

The model is one layer.

01 / Model

The commodity

The right model for the job, not the most expensive one by default. Routing, fallbacks and cost limits.

02 / Harness

The body

Deterministic tools, curated context, state and guardrails. The part the model does not write for you.

03 / Data

The ground truth

Documents, ledgers, market feeds and internal systems. Clean sources, explicit permissions, no magic.

04 / Evaluation

The proof

Task-level tests, traces and human review. We know what good looks like before the agent goes live.

05 / Run

The loop

Monitoring, cost control and a clean handoff. The agent reports what it tried, not just what it produced.

04 / From the blog

How we think about agents.

All writing ↗

The harness is the moat

Why the model is a commodity and the system around it is the advantage.

Read ↗

AI agents and prediction markets

How LLM forecasting is changing research and execution in event markets.

Read ↗

Why agents fail in production

The missing integration layer: connectors, testing, cost controls and observability.

Read ↗

Start a conversation

Bring us the workflow. We will find the agent.

Tell us what the team does by hand today. We will answer with a point of view, not a brochure.

alex@vegaforge.dev ↗

VegaForge / AI systems