AI Trading Infrastructure (Reference Stack)

Skeleton prod AI trading infrastructure for research/recruiting: FastAPI router, GPU/CPU backtesting, ingestion, monitoring, reproducible evidence. Not for live trading LOL.

                   ┌──────────────────────────┐
                   │     Data Ingestion       │
                   │  (OHLCV, News, Embeds)   │
                   └───────────┬──────────────┘
                               │
                               ▼
                ┌───────────────────────────────┐
                │   Preprocessing & Alignment   │
                │ (time-index, feature windows) │
                └───────────┬───────────────────┘
                            │
     ┌──────────────────────┼─────────────────────────┐
     │                      │                         │
     ▼                      ▼                         ▼
┌──────────────┐    ┌────────────────┐      ┌───────────────────┐
│ Sentiment     │    │ Technical      │      │ Contradiction     │
│ Agent         │    │ Agent          │      │ Agent             │
│ (FinBERT/LLM) │    │ (Indicators)   │      │ (semantic vs tech)│
└──────┬────────┘    └──────┬─────────┘      └────────┬──────────┘
       │                    │                        │
       └──────────┬─────────┴──────────┬──────────────┘
                  ▼                    ▼
           ┌──────────────────────────────────┐
           │ Fusion Agent (Multi-Modal Model) │
           │  attention over embeddings+tech  │
           └───────────────────┬──────────────┘
                               │
                               ▼
                    ┌────────────────────┐
                    │   Execution Agent  │
                    │ (router: cooldown, │
                    │ safety, throttling)│
                    └─────────┬──────────┘
                              │
                              ▼
                      LIVE DECISION

---

TL;DR

• Backtests (demo dataset): Sharpe 2.95, MaxDD −0.36%, Ann. return ~2% (walk-forward). Evidence: evidence/backtests/metrics_summary.json + charts.

• Router SLOs: p99 ≈ 18–20 ms @ ~14k rps under k6 load. Evidence: evidence/benchmarks/k6_http_*.json + Grafana PNGs; Prometheus TSDB snapshot in evidence/prometheus/.

• Signals ledger: Published in evidence/signals/ (demo + full). Recompute hit-rate/PnL from CSV.

This is a production-style reference stack with public artifacts. Use it to audit methodology & results. Do not connect to real brokers.

---

Quickstart

# python env
python -m venv .venv && source .venv/bin/activate
pip install -r requirements-unified.txt
pip install -e .

# deterministic backtest
python -m src.strategy.backtest_runner \
  --config config/backtest_config.yaml \
  --seed 1337 \
  --output out/backtest_results.parquet

python scripts/evaluate_backtest.py \
  --input out/backtest_results.parquet \
  --outdir out/eval_demo

Expected artifacts:

• out/eval_demo/metrics_summary.json (≈ evidence/backtests/metrics_summary.json)
• out/eval_demo/portfolio_equity.png, return_distribution.png

---

Run the stack + bench the router

# bring up infra (Prometheus, Grafana, router, etc.)
cp .env.example .env
docker compose -f infra/compose.yaml up -d --build

# health check
curl -sf http://localhost:8080/healthz && echo "router OK"

# k6 (machine-readable export)
k6 run --vus 200 --duration 60s \
  -e BASE_URL=http://localhost:8080 \
  --summary-export out/k6_http_ci.json \
  benchmarks/bench_http_k6.js

# Prometheus snapshot (requires admin API enabled in compose)
PROM=$(docker ps --filter "ancestor=prom/prometheus:latest" -q | head -n1)
docker exec "$PROM" wget -qO- "http://localhost:9090/api/v1/admin/tsdb/snapshot?skip_head=true"
SNAP=$(docker exec "$PROM" sh -lc "ls -1 /prometheus/snapshots | tail -n1" | tr -d '\r')
docker cp "$PROM":/prometheus/snapshots/$SNAP out/prom_snapshot

Optional PNG export (proves Grafana renderer & panel IDs):

GRAFANA_PASS=${GRAFANA_PASSWORD:-admin}
curl -s -u "admin:$GRAFANA_PASS" \
 "http://localhost:3000/render/d-solo/fusionalpha-performance/Performance?orgId=1&panelId=2&width=1600&height=900" \
 --output out/router_latency_p99.png

---

Reproducible evidence

• Backtests: evidence/backtests/
• Router perf: evidence/benchmarks/
• Prometheus snapshot: evidence/prometheus/
• Signals ledger: evidence/signals/

We pin seeds, versions, and configs. See docs/METHODOLOGY.md for walk-forward protocol, TC/slippage, risk caps, and leakage guards.

---

Architecture (high level)

• Router: FastAPI + micro-batching + Redis caching
• Backtesting: GBM/Monte-Carlo engine (CPU/GPU), risk & position sizing, walk-forward
• Ingestion: NATS JetStream (at-least-once), S3 Parquet writers
• Observability: Prometheus (metrics), Grafana (dashboards, PNG exports)
• Perf: k6 scripts for HTTP + WebSocket

clients → FastAPI Router → Strategy/Models
   │            │
   │            ├─ Prometheus (metrics) → Grafana (dashboards)
   └─ k6 load ──┘
data → NATS → S3 Parquet → Backtests → Evidence/

---

Methodology & Data

• Full protocol: docs/METHODOLOGY.md
• Data sources & license: docs/DATA_SOURCES.md
• Disclaimer: This repo is for research/recruiting. Not investment advice.

---

Dev notes

• CI: Performance smoke test builds the stack, runs k6 with --summary-export, uploads JSON and Prom snapshot artifacts.
• Strict guard: CI fails on any stray code placeholders.
• License: MIT (see LICENSE).

---