The original blueprint.

§ 1 · Project vision

An autonomous AI agent operating in crypto markets, with a public dashboard showing every decision in real time. The project combines three potential revenue sources: trading profits, web traffic monetization (ads and donations), and growing an audience interested in the AI + crypto space.

1.1 Philosophy

• Fun first — the project is a hobby, not a job
• Total transparency — every trade, every mistake, every profit is public
• Step by step — paper trading first, real money only after validation
• Minimum budget — near-zero infrastructure costs, reinvest profits
• Zero hype — no return promises, just honest documentation

1.2 Three revenue streams

§ 2 · Architecture — the three brains

The agent operates with three modules in a hierarchy: Sentinel commands, Trend Follower directs, Grid Bot executes.

2.1 Brain 1 — Grid Bot (the soldier)

Always on. Places buy and sell orders at regular intervals within a range. Profits from natural oscillations without predicting anything.

• Range: defined dynamically based on recent volatility
• Levels: 8-12 per token (to be calibrated in paper trading)
• Works best in: sideways market with high volatility
• Risk: price exits range and doesn't return

2.2 Brain 2 — Trend Follower (the tactician)

Analyzes technical indicators on 1h and 4h timeframes. When it detects a clear trend, it modifies the Grid's behavior.

• EMA 12/26: golden/death cross as primary signal
• RSI (14): overbought (>70) / oversold (<30) filter
• Volume: trend confirmation — trend + high volume = strong signal
• Bullish trend → shift grid up, increase exposure
• Bearish trend → tighten grid, reduce exposure
• No trend → let Grid operate normally

2.3 Brain 3 — AI Sentinel (the commander)

Uses an LLM to analyze news and social media. Can override the other brains in case of critical events.

• Monitors: CoinDesk, CoinTelegraph, CryptoPanic, Twitter/X, Reddit, exchange announcements
• Every 15-30 min: collects news and passes them to the LLM
• Output: risk score (1-10) + opportunity score (1-10)
• Risk > 7 → reduce exposure or freeze + Telegram alert
• Opportunity > 8 → temporarily increase exposure (with stop-loss)
• Risk 4-6 → no action

§ 3 · Trading strategies

3.1 Strategy A — Solid altcoins (80% of capital)

Grid + Trend Follower on validated tokens. Fundamental rule: NEVER sell at a loss.

Token admission criteria:

• Market cap top 50-100
• Existing project > 6 months, doxxed team, audited code
• Real daily volume, listed on 2+ major exchanges

3.2 Strategy B — Shitcoin pump (20% of capital)

Risk allocation for speculative trades. Losses accepted.

On CEX: limited to new listings (2-3/week). On Solana DEX: more opportunities but added complexity, to be implemented in phase 2.

§ 4 · Profit & loss management

4.1 Tiered take profit

4.2 Weekly rebalancing

• Every Monday: if a token > 30% of capital, trim the excess
• If a token < 5% and still valid, may reallocate

4.3 Cash management

• Crystallized profits: 50% reinvested in operational pool, 50% to stablecoin reserve
• Reserve only grows — never reinvested automatically
• Withdrawals only on explicit user request

4.4 Net P&L

No fee cap. The agent monitors net P&L (gains - fees). If daily P&L < -€10 on Strategy A, it stops. If the fee/profit ratio is too thin, it widens grid intervals.

§ 5 · Hardcoded rules

Implemented in code, not modifiable by the AI.

§ 6 · Public dashboard

The dashboard is the real product. It must be published from day one of paper trading to build an audience.

6.1 Structure

Homepage: Live portfolio (percentages, total P&L, chart from start). Status of three brains: Grid, Trend, Sentinel with risk score. Last trade with natural language explanation.

Operations page: Full history filterable by token, type, date. Each operation shows the decision reason.

Diary page: Daily agent commentary with defined personality and tone. Weekly analysis, posts on memory resets or strategy changes.

Info page: Project history, architecture, who's behind it.

Monetization: Banner ads (crypto sector: CPM €10-20/1000 views). Donations (Buy Me a Coffee, Ko-fi, crypto).

6.2 Tech stack

§ 7 · Learning system

The LLM doesn't learn in the classical sense. Intelligence grows through increasingly rich context fed to the model.

7.1 Structured memory

Database with complete history. Before each decision, the prompt includes: last N trades, latest Sentinel analyses, recent P&L, user feedback.

7.2 Auto-generated rules

A weekly job analyzes history and generates new rules. E.g.: "In the last 4 weeks, when RSI < 25 on TOKEN X, bounce within 6h." Saved in DB, loaded into prompt.

7.3 Daily diary

1 API call/day at end of day. The LLM writes a commentary with the agent's personality. Published on the dashboard. Cost: pennies.

7.4 Weekly report

1 API call/week. Analyzes trades, identifies patterns, generates rules. Can be done semi-manually (data via Telegram, rules generated in Claude Pro/Max chat). Cost: negligible.

7.5 Feedback loop

Telegram alerts with user response ("good decision" / "should have waited"). Feedback saved and used as future context.

§ 8 · Agent memory management

All memory lives in the database. Full control: reset, restore, compare versions.

Reset = clear rules table in DB. Click on admin dashboard (not public). Agent restarts with only hardcoded base rules.

§ 9 · Notifications (Telegram bot)

§ 10 · Cost analysis

10.1 Infrastructure

10.2 API optimization

The Sentinel accounts for 99% of API cost. Strategies:

• Use Haiku (or equivalent cheap model) instead of Sonnet/Opus
• Reduce frequency: 30-60 min normally, accelerate only in high volatility
• Prompt caching for the Sentinel system prompt
• Pre-filter news in Python before sending to LLM
• Evaluate local open-source models (Llama, Mistral) on Mac Mini for pre-filtering

§ 11 · 12-month simulation (conservative)

Assumptions: Str. A +2%/month net, Str. B volatile, profits 50/50 reinvest/reserve.

§ 12 · Development plan

Phase 0 — Setup (Week 1): Choose exchange (Binance/Bybit) and create KYC account. Configure API key (read + trade only, NO withdraw). Setup Python: ccxt, FastAPI, SQLite/Supabase. Create Telegram bot. Choose agent name and identity + register domain.

Phase 1 — Grid Bot + paper trading (Weeks 2-3): Implement Grid Bot in paper trading. Trade logging to database. Basic admin dashboard + Telegram notifications. Publish public dashboard (even basic).

Phase 2 — Trend Follower (Weeks 3-4): Calculate indicators (EMA, RSI, Volume). Trend Follower logic + Grid interaction. Parameter calibration on paper trading data.

Phase 3 — AI Sentinel (Weeks 4-5): Integrate news feeds (CryptoPanic, RSS). Sentinel prompt with LLM. Automatic daily diary. Sentinel override on other brains.

Phase 4 — Full dashboard (Weeks 5-6): Design with v0, React implementation, Vercel deploy, Supabase realtime.

Phase 5 — Extended paper trading (Weeks 6-10): Minimum 4 weeks. Optimization, fixes, first rule generation. Dashboard already live.

Phase 6 — Go live (After week 10): Only if paper trading is satisfactory. Start with €100-200, not €500. Scale gradually. Possible VPS migration.

§ 13 · References & resources

13.1 Open source projects

Freqtrade (main architectural reference) — Open source Python crypto bot. Backtesting, paper trading, WebUI, Telegram. Very similar stack to ours.

AI-Trader (dashboard + AI reference) — Live public dashboard with AI agents trading in real time. Copy trading, OpenClaw integration.

13.2 Market data & warnings

Failure statistics:

• 73% of automated trading accounts fail within 6 months
• 53.2% of all crypto projects failed (13.4M) — CoinGecko Jan 2026
• 11.6M tokens ceased in 2025 alone
• October 2025: crash — $19B liquidated in 24h

AI incidents:

• Lobstar Wild: decimal error, $441,000 lost in a single transaction
• GPT-5 test NOV1.ai: -62% capital in a few weeks
• May 2025: AI bots sell $2B in 3 minutes, amplifying a flash crash

Regulatory warnings:

• CFTC (USA): warning against AI bots with guaranteed returns
• MiFID II (EU): paid copy trading may require authorization

13.3 Tech stack

13.4 Conceptual inspiration

• Felix Craft AI (Nat Eliason) — AI agent "CEO" with public dashboard. Reference for storytelling and transparency.
• OpenClaw (Peter Steinberger) — autonomous open source AI agent. Reference for external tool integration.

§ 14 · Open decisions