LLMTokenStreamQuantEngine

Real-time LLM token stream ingestion → semantic signal extraction → risk-gated trade signal generation. Sub-microsecond latency. Live OpenAI streaming. Zero managed dependencies in the hot path.

---

What It Does

Connects directly to OpenAI's gpt-4o streaming API over a raw TLS socket, ingests the token stream token-by-token as it arrives, maps each token through a semantic weight dictionary (bullish/bearish/volatile/crash/surge/etc.), accumulates directional bias and volatility signals, and fires risk-gated trade signals — all within microseconds of the token hitting the wire.

gpt-4o (api.openai.com:443)
        |
        | TLS socket, chunked HTTP/1.1, SSE
        v
LLMStreamClient  (background thread, per-request reconnect)
        |
        | token text (normalized, whitespace-stripped)
        v
Deduplicator     (sliding TTL window, in-process)
        |
        v
LLMAdapter       (exact-match dictionary, SIMD aggregate path)
        |  SemanticWeight { sentiment, confidence, volatility, directional_bias }
        v
TradeSignalEngine  (bias accumulator, volatility accumulator, signal cooldown)
        |  TradeSignal { delta_bias_shift, volatility_adjustment, timestamp }
        v
RiskManager      (magnitude gate, rate gate, drawdown gate, position gate)
        |
        v
stdout  (aligned columns: TIME | BIAS | VOL | LATENCY | PASS/BLOCK)

---

Live Demo

The screenshot above is a real run against gpt-4o — no mocking, no replay.

-- STATS  TPS: 32  TOK: 265  AVG: 1856us  P99: 6121us  PRESS:0.16  BKOF:1.0x  DEDUP:10  SIG-PASS:239  BLOCK:29  [!] P99 > target

• 32 tokens/sec live from OpenAI
• 239 signals passed risk gate in ~8 seconds
• P99 6121us end-to-end token-to-signal (dominated by SSE parse + dictionary lookup, not the signal math)
• DEDUP:10 duplicate tokens dropped cleanly

---

Features

Feature	Detail
Live LLM streaming	Raw TLS socket to OpenAI — no libcurl, no Boost, zero managed I/O dependencies
OpenSSL TLS	Full certificate verification via Windows system ROOT store injection
Chunked transfer decoding	HTTP/1.1 Transfer-Encoding: chunked stripped in the read loop
SSE parsing	data: lines extracted, [DONE] sentinel handled, delta-scoped JSON parse
Token normalization	Leading/trailing whitespace stripped, lowercased before dictionary lookup — handles " Bullish" → "bullish"
Semantic dictionary	40+ tokens: fear, certainty, directional, volatility, neutral — all tunable
SIMD aggregation	SSE2 path for multi-token sequence weighting (map_sequence_simd)
Deduplication	Sliding TTL in-process dedup, configurable window
Risk manager	Magnitude, rate, drawdown, and position gates — each independently configurable
Latency controller	P50/P99/max tracking, Welford online variance for semantic pressure, backoff multiplier
Hot-reload config	config.yaml watched on a background thread; bias/vol sensitivity updates live
OMS adapter	Mock OMS with position state callbacks; REST OMS adapter for real order routing
Output sinks	CSV, JSON, and in-memory sinks — pluggable via OutputSink abstract base
--debug-raw mode	Dumps raw socket bytes to stderr for 3 seconds then exits — for protocol debugging
--no-color mode	Strips all ANSI codes, ASCII-only dividers — clean in any terminal encoding
1,491 tests	Unit, integration, property-based, and chaos/fault-injection coverage

---

Getting Started

Prerequisites

• Windows: MSVC 19.44+ (Visual Studio 2022 BuildTools)
• CMake 3.20+
• vcpkg with: spdlog, yaml-cpp, gtest, openssl

Build (Windows / MSVC)

# Clone
git clone https://github.com/Mattbusel/LLMTokenStreamQuantEngine
cd LLMTokenStreamQuantEngine

# Configure (point vcpkg toolchain at your install)
cmake -B build -DCMAKE_BUILD_TYPE=Release `
  -DCMAKE_TOOLCHAIN_FILE="C:/vcpkg/scripts/buildsystems/vcpkg.cmake" `
  -DLLMQUANT_TLS_ENABLED=ON

# Build
cmake --build build --config Release

Run — Simulator Mode (no API key needed)

cd build\Release
.\LLMTokenStreamQuantEngine.exe --no-color

Plays back a built-in token loop (crash, panic, bullish, breakout, ...) through the full signal pipeline.

Run — Live Stream Mode (OpenAI gpt-4o)

cd build\Release
.\LLMTokenStreamQuantEngine.exe --stream "sk-proj-YOUR_KEY_HERE" --no-color

Connects to api.openai.com:443, authenticates, streams a financial sentiment completion every 5 seconds, and fires live signals. Use --no-color in Windows PowerShell to avoid CP850 encoding artifacts.

Debug Raw Socket Output

.\LLMTokenStreamQuantEngine.exe --stream "sk-proj-YOUR_KEY_HERE" --debug-raw

Dumps every raw byte from the TLS socket to stderr for 3 seconds — useful for verifying chunked encoding, SSE framing, or diagnosing auth failures.

---

Configuration

config.yaml controls all runtime parameters and is hot-reloaded without restart:

trading:
  bias_sensitivity: 1.0        # Scalar on BIAS accumulator
  volatility_sensitivity: 1.0  # Scalar on VOL accumulator
  signal_decay_rate: 0.95      # Per-tick decay on accumulated signal
  signal_cooldown_us: 1000     # Min microseconds between signals

latency:
  target_latency_us: 10        # P99 budget (alert fires if exceeded)
  sample_window: 1000

pressure:
  max_ingestion_rate_tps: 10000
  high_pressure_threshold: 0.8
  max_backoff_multiplier: 5.0

semantic_weights:
  fear_multiplier: 1.2
  bullish_multiplier: 1.0
  bearish_multiplier: 1.2
  volatility_multiplier: 1.1

---

Token-to-Signal Mapping

Category	Tokens	Effect
Fear / Panic	crash panic collapse plunge dump rout	Strong negative BIAS, high VOL
Directional Bullish	bullish rally surge breakout soar moon buy	Positive BIAS
Directional Bearish	bearish breakdown selloff short sell	Negative BIAS
Volatility	volatile spike whipsaw choppy erratic swing	VOL spike, neutral BIAS
Certainty	inevitable guarantee confident certain assured	Confidence boost

All entries are in src/LLMAdapter.cpp::initialize_default_mappings() and can be extended at runtime via load_sentiment_dictionary().

---

Architecture Notes

• No exceptions in hot path — all error surfaces return Result-style values
• Single background thread per stream — reader loop owns its socket, reconnects on EOF
• Per-request TLS reconnect — OpenAI closes after [DONE]; client reopens cleanly
• SSL_CTX reused across reconnects — only per-connection SSL* is torn down
• Windows CA store injection — vcpkg OpenSSL has no CA bundle; system ROOT certs loaded via CertOpenSystemStore + d2i_X509 at startup

---

Tests

cmake --build build --config Release --target tests
cd build\Release
.\tests.exe

1,491 tests across unit, integration, property-based (proptest-style), and chaos/fault-injection suites. Full pipeline end-to-end covered including backpressure cascade, circuit breaker recovery, dedup races, and retry backoff timing.

---

License

MIT — see LICENSE.