2026-08-04 00:49:37 -07:00
2026-08-03 00:25:57 -07:00

K4LSH1_OPS

Multi-coin LLM-driven 15-minute binary options trading bot on Kalshi. Three independent bots trade DOGE, SOL, and ETH simultaneously using a two-tier LLM decision engine.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        K4LSH1_OPS FLEET                         │
├──────────┬──────────┬───────────────────────────────────────────┤
│  DOGE 🐶 │  SOL 🔮  │  ETH 💎  ← 3 independent processes        │
│  .154    │  .71     │  .189    ← proxy-backed Binance intel     │
│  doge.db │  sol.db  │  eth.db  ← isolated tick/decision DBs     │
└──────────┴──────────┴───────────────────────────────────────────┘
        │                              │
        ▼                              ▼
┌──────────────┐            ┌──────────────────┐
│  qwen3.5:4b  │  borderline│  ornith:latest   │
│  MacBook     │──────────▶ │  GamingPC .186   │
│  (0.4s gate) │  escalate  │  RTX 3070 (0.86s)│
└──────────────┘            └──────────────────┘
        │
        ▼
┌──────────────────────────────────────────────┐
│  Kalshi V2 API (api.elections.kalshi.com)    │
│  /portfolio/events/orders — bid/ask format   │
└──────────────────────────────────────────────┘

Data Inputs Per Decision

Every 5-minute decision cycle feeds the LLM a full market snapshot:

Source Data Frequency
Kraken Spot price (per coin) 20s tick
Binance (via proxy) 24h change, 24h range, 24h volume 5min cached
Kraken OHLC 1m/5m/15m/1h/2h multi-timeframe trend per cycle
SQLite ticks RSI(14) on 1m and 5m timeframes per cycle
Kalshi orderbook YES/NO ask, spread, depth imbalance per cycle
Fear&Greed Index Market sentiment (0-100) 5min cached
Mempool.space BTC fast fee (sat/vB) 5min cached
Orders DB Per-coin 24h win rate, recent performance per cycle
Volatility calc Market regime (choppy/trending/normal) per cycle

Decision Engine

Two-Tier LLM Voting

  1. Fast gateqwen3.5:4b on MacBook (0.4s). Handles clear signals. If confident (≥0.70) → trade immediately.
  2. Deep verifyornith:latest on GamingPC RTX 3070 (0.86s). Called when gate is borderline (<0.70 or unclear). Final verdict.

Safety Gates (checked before any order)

  • RSI contrarian — RSI > 85 → auto-fade DOWN, RSI < 15 → auto-fade UP (no LLM needed)
  • BTC macro correlation — Won't short alts when BTC 24h > +0.5% unless conf > 0.78
  • Adaptive confidence — Win rate adjusts minimum confidence (loses → tighter, wins → relaxed)
  • Self-adapting overrides — qwen analyzes last 20 resolved bets every 10 min, tunes thresholds
  • Daily loss cap — Hard stop at configurable daily loss limit
  • Market expiration guard — Verifies market open before placing order

Position Sizing

  • Max $1.00 per trade (configurable via max_spend_cents)
  • Confidence-scaled contracts: <0.70 = x1, 0.70-0.80 = x2, 0.80-0.90 = x3, >0.90 = x4
  • Auto-capped so contracts × price ≤ max_spend
  • Taker at ask when ≤ 55¢, maker at 50¢ when book is thin

Hedge Engine

Scans every cycle for open positions where buying the opposite side locks ≥8¢ profit. Auto-executes counter-orders.

Files

File Purpose
bot.py Main bot — decision engine, Kalshi client, order placement
dashboard.py Flask dashboard — fleet view at /fleet, single-coin at /
test_bot.py Unit tests — momentum, RSI, hedge, decision logic, LLM parsing
config.json Base config (shared defaults)
doge.json / sol.json / eth.json Per-coin configs with proxies, LLM URLs, DB paths

Quick Start

# Launch all 3 bots
cd ~/kalshi-bot
.venv/bin/python bot.py --config doge.json --lock doge &
.venv/bin/python bot.py --config sol.json --lock sol &
.venv/bin/python bot.py --config eth.json --lock eth &

# Launch dashboard
.venv/bin/python dashboard.py

# Run tests
.venv/bin/python test_bot.py

Dashboard

  • Fleet view: http://localhost:5053/fleet — all 3 coins, color-coded, PNL, decision trace
  • Single coin: http://localhost:5053 — controls, orderbook, momentum chart
  • API: GET /api/fleet — all coins JSON, GET /api/state — single coin state

Infrastructure

Component Host Purpose
MacBook localhost:11434 qwen3.5:4b fast gate
GamingPC 10.30.20.186:11434 ornith:latest deep verify (RTX 3070)
proxy1 10.30.20.154:3128 DOGE Binance intel (Brazil exit)
proxy2 10.30.20.71:3128 SOL Binance intel (EU exit)
proxy3 10.30.20.189:3128 ETH Binance intel (Brazil exit)
Gitea 10.30.20.149:3000 Code repo
Description
K4LSH1_OPS — multi-coin LLM bot: qwen fast-gate + ornith deep-verify, RSI contrarian, BTC macro gate, self-adapting thresholds, fleet dashboard
Readme 5.3 MiB
Languages
Python 99.5%
Shell 0.5%