← All projects

2026 · Autonomous trading intelligence system

ATIS

A near-institutional-grade multi-agent AI system for swing trading on Indian equities. 59 agents, a live Neo4j knowledge graph, a Rust hot path, and a 3-machine K3s cluster at home.

In Progress 2026

Why I built this

I wanted to replace intuition-based trading with something systematic and auditable, where every signal has a traceable, backtested reason. I also wanted to understand what near-institutional-grade tooling actually looks like when you strip away the Bloomberg terminal and the vendor data feeds. The constraint of free data only forced cleaner thinking: if the architecture works here, it works anywhere. The Rust hot path came from a real pain point (broker connection limits and Python's 5 to 15 ms latency floor), not from wanting to use Rust for its own sake.

59 Agents across 7 layers
<5μs Tick-to-signal latency (Rust)
85/100 System effectiveness score

Architecture

Six-tier pipeline

  1. 01

    Knowledge Pipeline

    Ingest research papers and filings, extract entities, build deduplicated theses with provenance.

    ingestionextractionthesis building
  2. 02

    Backtesting Engine

    Walk-forward and Monte Carlo validation with full transaction cost accounting on GPU.

    walk-forwardMonte Carloportfolio sim
  3. 03

    Daily Screening

    Regime detection, news ingestion, technical scanner, and fundamental filter run before market open.

    regimenewsscannerfundamentals
  4. 04

    Graph Reasoning

    GraphRAG multi-hop traversal, subgraph narration, and causal chain analysis over the knowledge graph.

    GraphRAGsubgraph narratorcausal chains
  5. 05

    Live Monitoring

    Real-time position monitor, Elo updater, risk checks, and drawdown protection on every tick.

    positionsElo updaterriskdrawdown
  6. 06

    Output

    Ranked daily signals, dummy trades for tracking, live dashboard, and Telegram approvals.

    ranked signalsdummy tradesdashboardTelegram

GraphRAG reasoning chain

Agents never receive raw Cypher output. Every reasoning step is verified against Neo4j; if hallucination rate goes above 15 percent, a Telegram alert fires.

Neo4j KG 600 stocks · theses GraphRAG multi-hop · max 3 Mistral 7B subgraph narrator 7-step CoT regime → trade Verification claims vs Neo4j Telegram alert hallucination > 15% fail

59-agent swarm, 7 layers

Ingestion (7), Knowledge (9), Analysis (5), Signal (7), Infrastructure (14), Validation (3), Additional (10). Each agent has a fixed LLM, temperature, and output contract.

Neo4j knowledge graph

  • Ontology · node and edge type schema
  • Knowledge · thesis registry with Elo scores
  • Market Structure · 600-stock dependency graph
  • Temporal · alpha decay over time
  • Event · historical causal chains
  • Decision · agent verdict audit trail

20 node types · 25 edge types

Rust hot path

Single Dhan WebSocket feeds Chronicle Queue. Rust fans out ticks to live, dummy, and backtest consumers from one connection. Under 5 μs end to end. Eliminates Python's GIL overhead and broker connection exhaustion.

3-machine K3s cluster

  • M1 · RTX 3080 Ti, 12 GB · orchestrator, Qwen2.5 14B, GPU backtesting
  • M2 · A1000, 6 GB · Mistral 7B, knowledge graph generation
  • M3 · T600, 4 GB · PostgreSQL, Neo4j, Qdrant, Redis

Ceph replication factor 2 · any one node can fail

Tech stack

Technologies used

core

Python 3.11Rust (Chronicle Queue hot path)LangGraph (agent orchestration)Neo4j (knowledge graph)Qwen2.5 14B + Mistral 7B via OllamaClaude API (validation, 5% of ops)

infra

K3s (3-node cluster)Ceph (distributed storage)PostgreSQL + streaming replicationRedis Streams + SentinelQdrant (vector DB)Docker

tools

VectorBT (GPU backtesting)Optuna (Bayesian optimisation)spaCy en_core_web_trf + custom NERscipy (correlation, regime clustering)CeleryStreamlitTelegram Bot API

Key highlights

Proof points

  1. 01

    59 agents across 7 layers covering knowledge ingestion, backtesting, daily screening, graph reasoning, live monitoring, and output.

  2. 02

    Neo4j knowledge graph with 6 layers, 20 node types, and 25 edge types. Enables second and third-order causal propagation across 600 NSE/BSE stocks.

  3. 03

    GraphRAG verification agent checks every LLM reasoning claim against Neo4j facts and logs hallucination rate per agent. Alert fires above 15 percent.

  4. 04

    Rust and Chronicle Queue hot path achieves under 5 μs tick-to-consumer latency and solves broker connection exhaustion with a single WebSocket.

  5. 05

    Walk-forward and Monte Carlo backtesting via VectorBT with full transaction cost accounting. System effectiveness: 85/100 on free data.

  6. 06

    Elo-based thesis lifecycle management and temporal edge decay mean the system's knowledge improves over time without manual intervention.

  7. 07

    3-machine home K3s cluster with Ceph replication. If any one machine fails, the other two hold all Tier 1 data and trading continues.

Focus areas

LangGraphNeo4jRustVectorBTGraphRAGK3sKafkaQdrant

Explore the work