ZEVIQ
Contact →

Platform

The research platform behind our agents.

Agents are only as good as the software around them. They need data to work from, memory to build on, tests to be measured against, and guardrails they cannot move. We build that platform in-house.

Platform · concept preview

Zeviq Studio: where agents do the research.

A look at the interface we're building for our agents and researchers: grounded research, hypotheses with their pass criteria, validation gates, agent activity and risk limits, in one place.

Concept preview: illustrative interface and synthetic data. The platform is in development.

ZEVIQ STUDIOresearch workspace

Research assistant

answers only from the research library

How should position size relate to the stop distance?

Size is the output, not the input: fix the dollar risk per trade first, then divide it by the distance to the stop[1]. A wider stop means a smaller position, so every trade risks the same amount[2]. Full Kelly sizing is a ceiling, not a target; fractional sizing gives up a little growth for much smaller drawdowns[3].

[1] Sizing notes · §Stop-based sizing[2] Risk handbook · p.140[3] Kelly notes · §Fractional Kelly
✓ citations verified 3/3relevance 0.93 ≥ 0.77sources: 2 authors agree

What does the Hindenburg Omen predict for next week?

Not in our knowledge base. No source in the library covers this; the closest passages scored 0.31 (threshold 0.77). The assistant won't answer from general knowledge.

declined: low relevancesuggest: add a source
  1. 1 ms query vector cache hit
  2. 17 ms vector + keyword search · 40 candidates
  3. 190 ms rerank · top score 0.93
  4. — draft answer from 6 passages
  5. — check every citation against its passage
  6. — label: [library] · [analysis] · [model]

Every step traced and timed. Answers that can't be grounded are declined.

The research assistant, knowledge system and evaluation tracing exist today in early form; the hypothesis board, validation reports, agent console and risk monitor are in development. All data shown is synthetic.

01 — Components

Six parts, one system.

What each part does, and where it stands today.

  • Knowledge system

    Retrieval over a curated research library, using hybrid vector and keyword search with reranking. Every source carries a trust tier, and drafts are never searchable.

    Live

  • Grounded answers

    Answers cite their sources, and the citations are checked in code. Below a relevance threshold, the system says “not in our knowledge base” instead of guessing.

    In development

  • Research agents

    Agents that turn ideas into hypothesis cards, with the mechanism, pass criteria, and falsification written before any test, plus independent reviewer agents that check the work.

    In development

  • Evaluation & tracing

    Independent answer keys with held-out splits. Every operation is timed (p50/p99) and traced, so quality, latency, and cost are measured rather than assumed.

    Live

  • Data & backtesting engine

    Point-in-time market data, walk-forward backtests with costs, and known-answer tests for the engine itself.

    In development

  • Execution & risk

    Deterministic order and risk code. Agents will be able to halt trading but never loosen a limit.

    Planned

02 — Architecture

From source to live, measured at every step.

  1. 01Sources
  2. 02Knowledge system
  3. 03Research agents
  4. 04Hypotheses
  5. 05Backtest & validation
  6. 06Paper
  7. 07Live
Evaluation & tracingspans every stage
Fig. 1 — How work moves through the platform (simplified). Ideas only advance when they pass the stage before.
Our research standardContact us