LISTENING · TELEGRAM + BALE DISTILBERT · TRAINED ON YOUR CHANNEL

Channel signals.
Soodchin execution.
Zero hesitation.

SOODCHIN watches your Telegram / Bale channels, understands BUY / SELL with a DistilBERT two-head classifier fine-tuned on your channel's own style, asks you to Accept / Deny — then fires the MT5 order and manages TP/SL live.

Run live pipeline demo Why channel training?

0sstale-message cutoff

0msDistilBERT inference

1slive TP/SL poll loop

24/7headless + SQLite

Honest note: best accuracy comes after fine-tuning on YOUR channel's history. Generic model works — trained model usually works far better. Nothing guaranteed in markets.

soodchin — live execution feed ● LIVE

GOLD SNIPER FXToday 14:32:07 · VIP

XAUUSD BUY NOW @ 2384.20
SL 2378.50 TP1 2390.00 TP2 2396.50
Risk 1%

DistilBERTBUY XAUUSD · conf 98.7%

Admin gate → 15s

ORDER#88412 BUY 0.50 XAUUSD

FILL2384.34 → TP1 50% + BE

MT5 BRIDGEConnected · 12ms

MODEApproval → Autonomous

TP MANAGER1s poll · partials

TELETHON + BALE API◆DISTILBERT TWO-HEAD◆MT5 PYTHON BRIDGE◆ADMIN APPROVAL GATE◆LIVE TP/SL MANAGER◆SQLITE · RSS NEWS

01 — THE PIPELINE · FULL SYSTEM

Three nodes. Six phases.

From channel post to broker fill — with dedup gates, AI verify, admin approval and live position management. This is the real Bale Bot architecture, simplified for showcase.

STAGE 01

Telegram / Bale

Source of truth

Telethon user-client + Bale HTTP polling. Stale check (120s), SHA256 dedup, rate-limited FIFO queue. Edits, forwards, replies handled.

→ AI

STAGE 02

Soodchin Brain

Channel-trained DistilBERT

Regex entities → DistilBERT two-head (is_signal? + action/symbol) → LLM verify → regime check → proposal. Exposure guard blocks pyramiding.

→ MT5

STAGE 03

MetaTrader 5

Hands that trade + guard

Spread / margin / algo-trading checks → dynamic lot → drift-adjusted SL/TP → order_send (IOC→FOK→RETURN) → 1-sec live TP loop.

PHASE 1

Signal Ingestion

Telethon listener → Gate 1 stale (>120s drop) → Gate 2 SHA256 dedup → ChannelSignalQueue FIFO. Survives restarts via persistent dedup store.

PHASE 2

Signal Processing · 10 steps

MT5 sync → state-aware dedup → regex entities → DistilBERT parse → exposure guard → event save → AI verify → SMA/ATR regime → confidence + trade plan.

PHASE 3

Admin Gate

Proposal card with inline buttons. Autonomous GOLD mode or 15-sec fallback auto-execute. Telegram failure → auto-execute per config.

PHASE 4

MT5 Execution

Symbol validate → spread < MAX → ATR SL fallback → risk% lot → TP side check → split tickets per TP → order_send with filling-mode loop.

PHASE 5

Live TP/SL Manager

1-second poll: partial close per tp_rules.json (TP1 50% → breakeven, TP2 → TP1…), SL trailing, manual-edit sync, close reports to admin.

PHASE 6+

Macro Chat + News

/chat agent with FTS5 hierarchy search + RSS forex_news grounding. /propose SQL cards. SQLite ×3: platform, hierarchy, news.

02 — INTERACTIVE SHOWCASE

Don't believe it. Test it.

Paste a signal like your channel would post it. Note: demo uses generic parsing — your channel-trained model would score higher.

INCOMING MESSAGE — click a sample or type your own

1Listen

2DistilBERT

3Approve+Price

4Execute

Low confidence (<0.85) or spam → held + logged, never forced. Exposure guard blocks same-symbol stacking.

SOODCHIN BRAIN OUTPUT— ms

{
  "is_signal": "—",
  "action": "—",
  "symbol": "—",
  "confidence": "—"
}

sysReady. Awaiting signal…

03 — CHANNEL TRAINING · READ THIS FIRST

Generic AI guesses. Trained AI knows your channel.

Every signal provider writes differently. For best accuracy the DistilBERT model is fine-tuned on YOUR channel's own message history first. It learns the style, then it works. Not guaranteed — but from what we know, it usually works.

How channel training works

1. Export history — we pull 500–2000 past messages from your channel.
2. Auto-label + verify — signals vs spam vs management commands, with human-in-the-loop inbox.
3. Fine-tune DistilBERT — two-head model trained on Colab T4 GPU on your style.
4. Deploy lightweight — quantized model drops into your bot. Offline inference ~38ms.

Generic model~64–72%

Channel-trained (same channel)~90–96%*

* Typical uplift observed in internal tests. Varies by channel consistency. Markets carry risk — accuracy is not a profit guarantee.

DistilBERT · two heads

models/khanbours_model/

Head 1 — Detection: is it a real signal? spam? management command?
Head 2 — Extraction: BUY/SELL, symbol, entry, SL, TP1–TP3 + confidence.
Fallbacks: regex parser + LLM verify (Lizh / OpenRouter / DeepSeek).

msg → DistilBERT → {is_signal: 0.99, action: BUY, …}

Skip training and the bot still runs — but expect misreads on slang, TP celebrations and forwarded formats. Training is strongly recommended.

04 — HOW DOES IT WORK?

From signal to broker fill.

Listen — Telegram / Bale tap

Persistent Telethon + Bale session. Stale >120s dropped, SHA256 dedup, FIFO queue. Spam and celebration posts filtered by Head 1.

live_trader.py · Telethon · 24/7

Understand — your channel's DistilBERT

Regex entities + channel-trained DistilBERT two-head: is_signal? → action, symbol, entry/SL/TP, confidence. Second LLM call verifies.

DistilBERT · ~38ms · offline

Decide — admin gate

Proposal card with Accept / Deny. Autonomous GOLD mode or 15-sec fallback. Exposure guard + SMA/ATR regime + risk score decide.

handlers/callbacks.py · pending_approvals

Execute — MT5 with guardrails

Spread < MAX, margin + algo-trading checks, ATR SL fallback, risk% lot, drift re-anchor, split tickets per TP, IOC→FOK→RETURN loop.

executor.py · order_send

Manage — 1-second TP loop

live_tp_manager polls every second: TP1 close 50% → SL to breakeven, TP2 → TP1… Manual phone edits synced, outcomes reported.

live_tp_manager.py · tp_rules.json

05 — THE DISTILBERT ENGINE

Small enough to forget.
Trained to know you.

A DistilBERT two-head classifier (KhanBours model), fine-tuned per channel on Colab T4. Runs locally — with regex + 3-engine LLM matrix (Lizh / OpenRouter / DeepSeek) as verify fallback.

DistilBERTtwo-head

38 msavg inference

$0.00local classify

Offlineprivate

	Soodchin DistilBERT	Cloud LLM wrapper
Speed	~38ms	1.5 – 4s
Channel style	Learns yours	Generic guesses
Cost / 1k signals	$0	$8 – $25
Works offline	Yes	No

Python 3.8+Telethon + Bale APIMT5 nativeSQLite ×3RSS + V2RayColab T4 train

live_trader.py — real flow

# Soodchin — channel-trained DistilBERT path
from models.khanbours_model import KhanBours
from executor import execute_trade

brain = KhanBours("models/khanbours_goldvip/")  # trained on YOUR channel
r = brain.parse(msg)  # Head1: is_signal 0.99 · Head2: BUY XAUUSD

if r.is_signal and r.confidence > 0.85:
    propose_admin(r)  # Accept / Deny · 15s fallback
    if approved: execute_trade(r)  # MT5 + tp_rules.json loop

Without training: same code, lower confidence — usually misreads slang.

06 — CAPABILITY MATRIX

Everything a manual copier wishes it could do.

Channel-trained accuracy

Fine-tuned on your provider's format. Spam vs signal vs commands separated.

Admin approval gate

Inline Accept/Deny buttons, 15s fallback, autonomous GOLD mode, Telegram-fail auto-execute.

MT5 executor

Spread filter, ATR SL fallback, risk% lots, drift re-anchor, split tickets, IOC→FOK loop.

Live TP manager

1-sec poll: partials per tp_rules.json, breakeven + trailing, manual-edit sync.

Dedup + exposure guard

SHA256 persistent dedup, state-aware dup, anti-pyramiding, regime filter.

SQLite + audit trail

platform_data, hierarchy FTS5, forex_news RSS. Every signal → ticket logged.

Macro chat agent

FTS5 + RSS-grounded market Q&A with session memory.

Private by design

Local DistilBERT inference. Channels, lots, broker stay on your Windows VPS.

07 — FORWARD TEST · BE UPDATED

We're still in the lab. Here's the honest status.

Demo showcase. Paper execution works — live P&L validation calibrating now. No fake profit curves.

PAPER FORWARD-TEST · XAUUSD + EURUSD

Equity curve — calibrating

COLLECTING TICKSFULL REPORT SOON

Results lockedForward test running · unlocks after 500 closed trades.Notify me

TRADES LOGGED312 / 500

TRAINED-CHANNEL ACC~93%*

AVG SLIPPAGE0.4 pips

UPTIME 30D99.9%

07 — PRICING · BETA

Free while we validate.

No paywall during the forward-test. Pick the seat that fits — paid tiers unlock only when live results publish.

PAPER BETA

Free

Full experimental console: channel training, 30 model tests, paper journal.

1 channel · 1 trained model
30 server model tests
Paper journal + dashboard

Request access

LIVE SEAT

$29/moafter validation

Real MT5 execution with the approval gate and 1-second TP/SL manager.

3 channels · 3 trained models
Unlimited parses · live execution
Priority training queue

Join the waitlist

DESK / FUND

Custom

Multi-seat, prop-firm drawdown sync, private VPS deploy, SLA.

Unlimited channels + seats
Dedicated training pipeline
Direct support line

Talk to us

Prices are beta placeholders and activate only after the 500-trade validation publishes. Paper Beta stays free through beta.

09 — ROADMAP

Where we are. Where it goes.

SHIPPED

Telegram + Bale listener + dedup
DistilBERT KhanBours v2
Admin Accept/Deny gate + 15s fallback
MT5 bridge + 1s TP loop

IN PROGRESS

500-trade forward validation
Per-channel training pipeline
Image-signal OCR
Public stats page

Prop-firm drawdown sync
One-click Windows VPS deploy
Web dashboard (SQLite → REST)
Strategy marketplace

10 — FAQ

Straight answers.

It will run without training — but for best accuracy you should fine-tune on your channel's history first. Each provider formats signals differently, and the DistilBERT model needs to learn that style. Untrained = more misreads. Trained = usually works very well. Not guaranteed — markets never are — but training is the difference.

Yes — a DistilBERT two-head classifier (KhanBours model): Head 1 decides is_signal / spam / management command; Head 2 extracts BUY/SELL, symbol, entry, SL, TP1–TP3 + confidence. Trained on Colab T4, deployed quantized for ~38ms offline inference. Regex + LLM matrix verify as backup.

Your choice. Approval mode sends Accept / Deny buttons with 15-second fallback auto-execute. Autonomous mode (e.g. GOLD) executes instantly. Telegram delivery failure can auto-execute per config.

A 1-second loop watches open positions: TP1 closes 50% and moves SL to breakeven, TP2 trails to TP1, etc. (tp_rules.json). Manual SL/TP edits from your phone are synced. Closures are reported to admin.

Windows VPS + MT5 + Python 3.8. SQLite needs no server. Your channel history for training (500–2000 messages ideal). That's it.

312/500 paper trades in. Full statement, win-rate, drawdown and slippage — good or bad — after validation. Request access to get notified.

10 — CONTACT

Talk to a human.

Support, channel-training slots, desk pricing — fastest on Telegram.

EARLY ACCESS · FOUNDING 50

Be first when the results drop.

Request experimental access — verify your email, join the queue, get a channel-training slot. One message, no spam.

Request access Log in

Free paper trial·Channel training included·10-min VPS setup

Channel signals.Soodchin execution.Zero hesitation.