Babingepet 2.0 — Architecture Document

01 — Architecture

System Architecture

Complete component map showing all four engines, shared infrastructure, AI layers, and data flows.

02 — Engine Deep Dive

BTC/USDT Engine

60-second cycle, 27 signals, 8-agent AI swarm, 12-source weighted voting decision engine.

Signal Pipeline

Binance API

60s pull

collector.py

OHLCV · OB · Trades

signals.py

VPIN · OFI · Edge

ai_estimator.py

8 agents · 10min

decision_engine.py

12-source vote

Simulator

Kelly · Trailing SL

12-Source Voting Weights

ml_xgboost

0.15

whale_flow

0.12

vpin

0.11

ai_consensus

0.10

mtf_alignment

0.10

regime

0.09

ofi_contrarian

0.08

vwap

0.07

funding_rate

0.06

as_fair_value

0.05

sentiment

0.04

hour_seasonality

0.03

AI Swarm Configuration

Consensus: 40% generalist (accuracy-weighted) + 60% swarm (equal weight)

Agent	Persona	Model
gemini	Generalist	gemini-2.5-pro
claude	Generalist	deepseek-r1:free
gpt	Generalist	qwen3.6-plus:free
momentum	Trend follower	gemini-flash
contrarian_shark	Mean reversion	gemini-flash
flow_reader	Microstructure	gemini-flash
macro_sage	Macro context	gemini-flash
conservative_filter	Risk manager	gemini-flash

Verdict Thresholds & Hour Logic

STRONG_BUY

≥ 0.28

BUY

≥ 0.20

SELL

≤ -0.20

STRONG_SELL

≤ -0.28

BLOCKED hours (UTC)

171819230

BOOSTED hours (UTC) — 1.3× edge

23789

03 — Engine Deep Dive

Altcoin Engine

4-hour cycle, 15-coin universe, 8-factor composite scoring, 5-agent AI swarm with sector gating.

8-Factor Composite Score

Momentum

0.28

Funding

0.13

BTC Alpha

0.13

Volume

0.12

Volatility

0.10

RSI

0.10

Orderbook

0.09

OI Proxy

0.05

AI Swarm Personas

ALPHA HUNTER

Seeks undervalued with momentum breakout potential, sector rotation plays

FLOW READER

OFI, funding dynamics, OI divergence, whale accumulation signals

MOMENTUM WOLF

Pure trend following, volume confirmation, breakout entries

MACRO CRYPTO

BTC correlation, sector narrative, macro catalyst alignment

RISK FILTER

Eliminates overleveraged, high-funding, overextended setups

        Vote scoring: Rank 1 = 3pts, Rank 2 = 2pts, Rank 3 = 1pt × conviction multiplier (HIGH=1.5×)
      

Coin Universe — 15 Active

ETHSOLBNBADAAVAXDOTNEARAPTSUIARBOPAAVEUNILINKINJ

BUY thresholdscore > 0.30

SELL thresholdscore < -0.20

Max positions3 concurrent

Sector limit2 per sector

Max hold24 hours

Fee accounting0.04% taker

Trailing stop+3% → 25% give

TP targetprice + 2.5×ATR

SL targetprice − 1.5×ATR

04 — Engine Deep Dive

IHSG Engine

Indonesian Stock Exchange coverage. 40-stock universe, 2× daily sessions, intraday overlay, news validation.

Session Schedule (WIB / UTC)

Session	WIB	UTC	Focus
Session 1 — Opening	09:00–12:00	02:00–05:00	Momentum + Volume surge
Session 2 — Afternoon	13:30–15:49	06:30–08:49	Mean-reversion + VWAP

        Hermes pre-session agent fires at 01:20 UTC (S1) and 05:50 UTC (S2) for news research.
      

Multi-Factor Model

Momentum1D · 5D · 20D ROC + accel

RSI14-period

VolumeSurge vs 20D avg

VolatilityGarman-Klass annualized

VWAP deviationAbove = bullish bonus

Range positionvs 52-week H/L

Sector strengthRelative vs benchmark

BUY threshold> 0.35

SELL threshold< -0.25

Kelly sizingHalf-Kelly · max 15%

40-Stock Universe

BBCABBRIBMRIBBNITLKMASIIUNVRICBPADROPTBAINDFKLBFSMGRBRISGOTOEMTKMNCNEXCLISATANTMMDKAMEDCPGASACESAMRTCPINJPFAMAPIERAABRPTTBIGTOWRSIDOHMSPGGRMMYORINTPWIKAPTPPSRIL

05 — Scanner

IHSG Scanner

On-demand bandarmology analysis. 5-signal composite, 51-broker intelligence database, whale-adjusted scoring.

5-Signal Weights

Acc/Dist

0.30

Bandar Cycle

0.20

Foreign Flow

0.18

RRG Rotation

0.17

Frequency (FQ)

0.15

51-Broker Intelligence

Tier	Count	Examples	Whale?
FOREIGN_BIG	10	RX, GR, CL, BK	2×
FOREIGN_MID	10	MQ, LG, FS	2×
DOMESTIC_BIG	10	CC, IF, NF, AI	2×
DOMESTIC_MID	13	YP, YU, XA	—
RETAIL	8	GG, SB, BI	—

3+ whale consensus → vote boost. Whale lots weighted 2× in net ratio.

Bandar Cycle Phases

①

ACCUMULATION

Sideways/down · broker buying · low vol

②

MARK UP

Rising price · vol surge · breakout

③

DISTRIBUTION

Sideways after markup · not accumulating

④

MARK DOWN

Declining price · supply flooding

⑤

EXIT

Declining price · very low volume

06 — Intelligence

AI Model Map

Complete mapping of AI models to functions, cost tiers, and fallback chains.

Model	Router	Used For	Cost Tier	Fallback
google/gemini-2.5-pro	OpenRouter	BTC generalist · Hermes analytics agents	PAID	gemini-flash
google/gemini-2.5-flash	OpenRouter	BTC swarm personas · Hermes Sentinel · IHSG/Altcoin swarm	LOW	deepseek-r1
deepseek/deepseek-r1:free	OpenRouter	BTC generalist "claude" slot	FREE	qwen3-plus
qwen/qwen3.6-plus:free	OpenRouter	BTC generalist "gpt" slot	FREE	deepseek-r1
Various OpenRouter	OpenRouter	IHSG 4-agent swarm · Altcoin 5-agent swarm	LOW	gemini-flash
XGBoost (local)	sklearn	BTC ml_xgboost signal (0.15 weight)	FREE	edge_estimate

Model Weight Logic

Generalist weights are dynamic — derived from last 50 predictions' accuracy. Weights are NOT normalized, preserving performance differences between models. Higher historical accuracy = higher weight on next prediction.

        Track record stored in crypto_ai_predictions — accuracy injected into prompts so models know their own hit rate.
      

Hermes Sentinel

On STRONG_BUY or STRONG_SELL (non-vetoed), spawns daemon thread → calls Gemini Flash with last 8 BTC memory cycles.

CONFIRM

CAUTION

WARN

Result written as hermes_note to btc_decisions table

07 — Agent Framework

Hermes Integration

Self-improving AI agent framework with SQLite persistent memory, 4 analytics agents, 6 Babingepet tools, and MCP bridge.

4 Analytics Agents (APScheduler)

Agent	Schedule	Goal
Signal Calibration	every 6h	Audit 50 predictions + 30 decisions
IHSG Pre-Session S1	01:20 UTC	Macro + per-stock news research
IHSG Pre-Session S2	05:50 UTC	Afternoon session stance
Altcoin Narrative	every 4h+10min	Narrative/catalyst per coin
Weekly Review	Sun 13:00 UTC	Full portfolio + predictions report

6 Babingepet Tools (Hermes)

TOOLbabingepet_btc_signalsLive BTC signal state from Supabase

TOOLbabingepet_portfolioCurrent positions + P&L

TOOLbabingepet_predictionsAI swarm p_up history

TOOLbabingepet_decisionsDecision verdicts + hermes_note

TOOLbabingepet_ihsg_picksSession picks + AI thesis

TOOLbabingepet_altcoin_picksTop 3 altcoin picks per cycle

MCP Bridge (7 read-only endpoints)

GET/hermes/statusSystem health + agent states

GET/hermes/btc_stateCurrent BTC signals + verdict

GET/hermes/portfolioAll open positions

GET/hermes/predictionsAI prediction history

GET/hermes/memoryHermes SQLite persistent memory

GET/hermes/ihsgIHSG session picks + AI analysis

GET/hermes/altcoinsAltcoin top 3 + swarm votes

      SOUL.md persona: For BTC/portfolio/IHSG/altcoin queries → call babingepet_* tools FIRST. Never web search for portfolio data. Concise, data-driven, quant analyst style.
    

08 — Risk Management

Risk Management Layer

All veto conditions, gates, and guards across all four engines.

BTC Engine — Veto Conditions

VETOAI agreement < 0.40 AND score > BUY threshold

VETORANGING regime vetoes most entries

GATEHour blocking: UTC 17,18,19,23,0 → no signals

GATEMin 5 of 12 sources must agree direction

GUARDSignal debounce: opposing signals → NEUTRAL

GUARDVWAP failed reclaim → score penalty (not override)

GUARDHermes WARN → flag in btc_decisions (advisory only)

Altcoin + IHSG Gates

VETOAltcoin: No new LONGs when BTC regime = TRENDING_DOWN

GATEAltcoin LONG requires score ≥ 0.05 at entry

GATEMax 2 coins per sector (L1/L2/DeFi/Meme/AI/Payment/Storage)

GATEMax 3 concurrent altcoin positions

GUARDAltcoin: score_warning flag when negative score for LONG

GUARDWilson CI hi < 0.45 → skip pattern (learning system)

GUARDIHSG news contradiction → swap pick with alternate

Kelly Criterion Position Sizing

Formula

f = (bp - q) / b
b = TP/SL ratio
p = win probability
q = 1 - p

BTC Parameters

Take Profit1.8%

Stop Loss0.6%

b ratio3.0×

ATR-Based (Alt/IHSG)

TPprice + 2.5×ATR

SLprice − 1.5×ATR

IHSG KellyHalf-Kelly · max 15%

09 — Adaptive Systems

Learning Feedback Loops

How trade outcomes feed back into future decisions across all three engines.

BTC Learning Loop

Trade Outcome

Win/Loss · PnL

Supabase Log

btc_sim_positions

Pattern Analysis

Hermes 6h calibration

learning_insights

Accuracy by hour/regime

Decision Engine

Modified sizing

Altcoin Learning (Wilson CI)

Condition	Action
ci_hi < 0.45	skip pattern
ci_hi < 0.50	reduce_size
ci_lo > 0.55	boost 1.2×
ci_lo > 0.65	boost 1.4×
ci_lo > 0.75	boost 1.6×

Min 20 trades for ACTIVE status. Patterns: ticker · sector · RSI regime · funding regime · score range · volume level

IHSG Learning (Bayesian)

Threshold	Status
8+ samples	Generated
10+ samples	ACTIVE
10+ samples	Applied to scoring

Empirical win rate uses Bayesian shrinkage with 50% neutral prior. Pattern key format: insight_type:condition_key (e.g. confidence:HIGH, session:Opening)

Pattern dimensions:

tickerconfidencesessionscore_bucketRSI_regimeaction_type

10 — Data Layer

Database Schema

Complete Supabase/PostgreSQL schema — all 16 tables with key columns and relationships.

BTC Tables

crypto_candles

idbigserial PKauto

timestamptimestamptzcandle open time

timeframevarchar1m · 5m · 15m

open/high/low/closenumericOHLCV

volumenumericquote volume

crypto_signals

idbigserial PK

timestamptimestamptz

vpinnumeric0–1 toxicity

ofinumericnormalized

edge_estimatenumericcomposite edge

signal_strengthvarcharSTRONG_BUY…STRONG_SELL

kelly_fractionnumericposition size

regimevarcharTRENDING_UP etc

crypto_ai_predictions

idbigserial PK

timestamptimestamptz

p_up_gemininumeric0–1

p_up_claudenumericdeepseek slot

p_up_gptnumericqwen slot

p_up_[persona]numeric×5 swarm agents

consensus_p_upnumeric40/60 blend

agreementnumericbull/bear ratio

model_weightsjsonbdynamic accuracy

btc_decisions

idbigserial PK

timestamptimestamptz

verdictvarcharSTRONG_BUY…STRONG_SELL

scorenumericweighted vote total

source_votesjsonball 12 source signals

vetobooleanvetoed flag

hermes_notevarcharCONFIRM/CAUTION/WARN

reasoningtextfull reasoning blob

crypto_trades

idbigserial PK

timestamptimestamptz

pricenumeric

quantitynumeric

is_whaleboolean>$150K trade

directionvarcharBUY · SELL

Portfolio + Market Tables

btc_sim_positions

idbigserial PK

sidevarcharLONG · SHORT

entry_pricenumeric

size_usdnumericKelly-sized

tp/slnumeric1.8% / 0.6%

trailing_highnumericpeak for trailing SL

statusvarcharOPEN · CLOSED

pnlnumericrealized

exit_reasonvarcharTP · SL · TRAIL · TIME

sim_portfolio

idbigserial PK

timestamptimestamptz

btc_equitynumericstarts $1,000

alt_equitynumericstarts $10,000

win_countinteger

loss_countinteger

altcoin_picks_history

idbigserial PK

timestamptimestamptzcycle time

picksjsonbtop 3 with scores

swarm_votesjsonbper-agent ranking

consensus_thesistextmerged narrative

btc_regimevarcharat time of picks

altcoin_sim_positions

idbigserial PK

symbolvarchar

sectorvarcharL1/L2/DeFi/etc

scorenumericentry score

size_pctnumericvol-scaled 12–25%

tp/slnumericATR-based

max_hold_attimestamptzforced close time

ihsg_picks_history

idbigserial PK

sessionvarcharOpening · Afternoon

picksjsonbtop 3 with scores

ai_thesistextper-ticker thesis

ai_overridebooleanAI changed quant order

news_validatedbooleannews check passed

learning_insights

idbigserial PK

enginevarcharbtc · altcoin · ihsg

pattern_keyvarcharinsight_type:condition

win_ratenumericWilson CI estimate

ci_lo / ci_hinumericconfidence bounds

sample_countinteger

actionvarcharskip/reduce/boost

statusvarcharACTIVE · PENDING

news

idbigserial PK

keywordvarchar5 BTC keywords

titletext

published_attimestamptz

sourcevarcharGoogle News RSS

11 — API Surface

API Endpoint Map

All FastAPI routes — method, path, purpose, and response shape.

BTC Endpoints

GET/BTC dashboard HTML

GET/api/signals{vpin, ofi, edge, signal_strength, kelly}

GET/api/predictions{p_up per model, consensus, agreement}

GET/api/decisions{verdict, score, source_votes, hermes_note}

GET/api/portfolio{equity, positions, win_rate, P&L}

GET/api/candles[{timestamp, ohlcv}] — 1m/5m/15m

GET/api/whale_trades[{price, qty, direction, is_whale}]

GET/api/news[{title, keyword, published_at}]

GET/api/regime{regime, confidence, regime_scores}

Market + Infrastructure Endpoints

GET/ihsgIHSG dashboard HTML

GET/api/ihsg/picks{session, picks[3], ai_thesis, scores}

GET/api/ihsg/learning{insights, pattern_key, action, ci}

GET/altcoinsAltcoin dashboard HTML

GET/api/altcoins/picks{picks[3], swarm_votes, consensus_thesis}

GET/api/altcoins/positions{open_positions, sector_breakdown}

GET/scannerIHSG Scanner HTML

POST/api/scanner/analyzebody: {ticker} → {5 signals, verdict, broker_pos}

GET/api/health{scheduler_jobs, last_run, db_status}

GET/hermes/*7× MCP bridge (read-only)

12 — Scheduling

Scheduler Timeline

24-hour visual timeline of all APScheduler jobs — cron and interval triggers.

UTC 00:00 — 23:59

000306091215182124

BTC Collector

every 60s — all day

BTC Signals

every 60s

BTC AI Swarm

every 10min

Altcoin Engine

00:00

06:00

12:00

18:00

Alt Price Check

every 15min

IHSG Session 1

02:00 UTC

IHSG Session 2

06:30 UTC

IHSG Price Upd

every 30min (market hours)

Hermes Pre-S1

01:20

Hermes Pre-S2

05:50

Signal Calibration

00:00

06:00

12:00

18:00

BTC Engine

Altcoin / Hermes

IHSG Engine

Calibration

Complete Job Registry

Job ID	Trigger	Interval	Purpose
run_collect	interval	60s	BTC data collection from Binance
run_signals	interval	60s	VPIN, OFI, Edge computation
run_sim	interval	60s	BTC paper trading simulator
run_ai	interval	10min	8-agent AI swarm prediction
run_ihsg	cron	02:00 + 06:30 UTC	IHSG session analysis
run_ihsg_price_update	interval	30min	Intraday IHSG price overlay
run_altcoins	interval	4h	Altcoin engine full cycle
run_altcoin_price_check	interval	15min	Open position P&L check
hermes_signal_calibration	interval	6h	Audit 50 predictions + 30 decisions
hermes_ihsg_s1	cron	01:20 UTC	Pre-session 1 research
hermes_ihsg_s2	cron	05:50 UTC	Pre-session 2 research
hermes_altcoin_narrative	interval	4h +10min offset	Altcoin narrative/catalyst research
hermes_weekly_review	cron	Sun 13:00 UTC	Full portfolio + predictions report

13 — Infrastructure

Deployment Architecture

VPS, Nginx, Systemd, process supervision — complete deployment topology.

Server Stack

OSUbuntu 22.04+ LTS

Domainmauuang.com

Web serverNginx (reverse proxy)

SSLLet's Encrypt

App serverUvicorn (port 8420)

FrameworkFastAPI + APScheduler

LanguagePython 3.12

DatabaseSupabase / PostgreSQL

Hermes memorySQLite (~/.hermes/babingepet.db)

Process Management

# Babingepet API
nohup uvicorn crypto.api:app \
  --host 0.0.0.0 \
  --port 8420 &

# Hermes Gateway (systemd)
systemctl enable hermes-gateway.service
systemctl start hermes-gateway.service

# hermes-gateway.service
[Service]
Restart=always
RestartSec=5s
ExecStart=/path/hermes-gateway
      

Deployment Diagram

14 — Data Flow

End-to-End Data Flow

From exchange API to learning loop — complete data lifecycle across BTC engine.

1

Exchange API Collection

Binance Futures → 1m candles, orderbook (depth 50), recent trades (1000), 24h ticker, funding rate, OI, L/S ratio, liquidations, whale detection (>$150K)

60s

2

Signal Computation

VPIN calculation (3 regimes), OFI normalization, Edge estimate (8 components), Kelly fraction, MTF alignment (5m+15m), regime classification

60s

3

AI Swarm Prediction

3 generalists × accuracy-weighted + 5 personas × equal weight → consensus p_up. Agreement score. Hermes Sentinel on STRONG signals.

10min

4

Decision Engine

12-source weighted vote → verdict. Apply veto conditions (regime, agreement, hour block). Min 5/12 source agreement. Threshold comparison.

60s

5

Paper Trading Execution

Kelly-sized entry. Trailing stop (+3% activation, 25% give). TP=1.8%, SL=0.6%. Max hold configurable. Write to btc_sim_positions.

60s

♻

Learning Feedback Loop

Trade outcome → Hermes calibration (6h) → learning_insights → pattern-based size adjustments → future decisions. Model accuracy weights updated from last 50 predictions.

6h cycle

Babingepet2.0

System Architecture

BTC/USDT Engine

Altcoin Engine

IHSG Engine

IHSG Scanner

AI Model Map

Hermes Integration

Risk Management Layer

Learning Feedback Loops

Database Schema

API Endpoint Map

Scheduler Timeline

Deployment Architecture

End-to-End Data Flow

Babingepet
2.0