Being built in AnkEDGE
Quant trading
Data processing, decision making, strategy ordering, adjustment linking, trade placement and hedging as agents, with risk limits in code and results feeding the next run.
The trading agent graphAgentic AI · LangGraph · Data to Data
Organise data and decisions
as a graph of AI agents.
We build multi-agent systems on LangGraph, the open source framework for agents that share one state. Each agent owns one decision, rules that must not bend stay in code, people approve what matters, and every outcome feeds the next run. We are building our own options trading platform, AnkEDGE, this way.
What LangGraph is.
An open source library, from the team behind LangChain, for building applications as a graph of steps. It is not a model and not a strategy. It is the control flow around models and code, available for Python and JavaScript.
- Shared state
- One record every agent reads and updates: the data, the analysis, the decision and who approved it.
- Nodes
- Each node does one job. Some call a language model, others run a statistical model or plain code.
- Conditional edges
- The graph chooses its own path from what is in the state, such as more evidence or a decision.
- Cycles
- Agents can loop back and try again, with a limit, instead of guessing.
- Checkpoints
- The state is saved after every step, so runs resume after failure and can be replayed for review.
- Interrupts
- A run pauses for a person and continues from the checkpoint when they answer.
Quant trading · in development
AnkEDGE: a trading desk built as an agent graph.
AnkEDGE's decision layer is being built on LangGraph. Data processing, decision making, strategy ordering, adjustment linking, trade placement and hedging are agents that share one state. Risk limits sit between them as plain code, and an order that breaks a limit goes back to be reordered rather than forced through.
Fills and position behaviour flow back into the next run, so the desk learns from its own trades. Strategies still pass backtesting and simulation before they reach the graph.
Prediction markets · capability
A price is a probability. The work is checking it.
An agent reads the resolution rules first. Three agents gather news, data and related market prices in parallel. Their evidence becomes one probability with reasons, and plain code compares it with the price after fees.
Usually there is no edge, and the graph logs that. When there is a gap, an analyst reviews the reasoning. Resolved outcomes score every estimate, so overconfidence shows up early. Analysis and review, not a promise of returns.
Prediction market analysis with agents
One pattern for large enterprises.
The roles stay the same in every industry: sense, analyse, decide, check, act, learn. The agents inside each role change. The check is always a rule in code or a person.
Capability
Prediction markets
Resolution rules read first, evidence gathered in parallel, probability compared with the price after fees, analyst review on any gap, and calibration on resolved outcomes.
Prediction market analysisCapability
Sports betting
Price and event feeds in, liability by market analysed, a reprice or suspension proposed, trader limits checked before anything is published.
Live exposure and position analyticsCapability
Casino and iGaming
Session and wallet events read for bonus abuse and safer gambling signals, an intervention chosen and checked against the operator's compliance rules.
Player risk cohorts in one consoleCapability
Insurance
Claims and documents read, fraud cohort signals checked, a fast track or referral proposed, and an adjuster approving anything outside the rules.
Insurance capabilityCapability
Mining
Sensor telemetry read for anomalies and wear, a maintenance plan proposed, a supervisor approving the work order, and the result compared with what failed.
Sensor telemetry and condition monitoring"Capability" means a pattern we build, linked to the delivered work it rests on. It is not a claim of a finished project in that industry.
Rules we build into every graph.
- Limits live in code
- No prompt can raise a limit or skip a check. Rules are versioned and released like any other code.
- Proposals, then actions
- Model output becomes an action only through the graph's rules or a person's approval.
- Language models off the fast path
- Models read and reason. Time-critical steps such as order placement run on code.
- Every run can be replayed
- Checkpoints and logs show which agent decided what, from which data, under which rules.
Data ready for agents.
Agents are only as good as the state they read. The foundations are the work we already deliver: event streams, point in time data, real-time reporting and triggers that turn live events into actions.
- Real-time market data processing
- Real-time analytics with Apache Pinot
- The Event Trigger Service
- Keeping production models right
Typical stack: LangGraph and LangChain, Python, Kafka, Apache Pinot, Elasticsearch and Kubernetes, with Prometheus and Grafana for monitoring. Models are chosen per node, hosted or open weight.
How it starts.
-
01
Decision workshop
We map the decisions you want organised, who owns each one, the data each needs and the rules that must never bend.
-
02
One graph in simulation
A small graph runs on your data alongside the current process, so its decisions can be compared before it acts.
-
03
Production, then the next graph
The graph goes live with monitoring and replay, our team runs it 24x7, and the next decision is added.
Tell us which decisions you want organised.
The decision, the data it depends on, who owns it today and the rules it must follow. A short list is plenty.
Your enquiry goes directly to Kartik Modi, our Managing Director, and an engineer replies by email.