Two engineers. No external funding. Built in full production with live users, real payment processing, and a public track record. No mockups. No waitlists.
Designed and built the entire AEGIS signal engine: XGBoost training pipeline, meta-labeling architecture, 8-layer signal filter, edge-score quality gate, regime-adaptive thresholds, beginner onboarding layer, FastAPI backend, Firestore integration, and WebSocket real-time delivery. Full-stack from ML to frontend.
Product direction, market positioning, user acquisition strategy, and go-to-market execution. Responsible for the subscription model design, pricing strategy, community building, and India-first user growth. Manages the business side so the engine can focus on signal quality.
Based in Dehradun, Uttarakhand, India · Operated under AEGIS v1.0
Retail crypto signal tools are dominated by two dysfunctional models: Telegram channels where humans manually post calls with 30–90 second delays and zero risk management, and simple indicator bots that fire on a single RSI or MACD crossing with no market context. Both are optimized for engagement, not for your money.
Every metric below is live and verifiable. The track record is public, unedited, and contains both wins and losses. No cherry-picking. No simulated results.
Three structural forces are converging in India that didn’t exist 3 years ago and won’t stay open for long.
AEGIS is a quantitative signal engine built on XGBoost with meta-labeling — a technique used by institutional quant funds to filter raw directional predictions into high-precision, tradeable signals.
Each of 63 monitored tokens has its own XGBoost classifier trained on ~7,000 hours of OHLCV data + 100+ engineered features: ATR, RSI, MACD, Supertrend, Fibonacci, support/resistance, funding rate, open interest, Fear & Greed index, BTC correlation, candlestick patterns, and more. Training uses purged time-series cross-validation — no lookahead leakage, strict temporal ordering with embargo windows.
After initial training, SHAP values rank features by cumulative importance. Features accounting for less than 10% of total importance are dropped, reducing from 100+ to 45–75 per model. Each token retains a different feature subset — EGLD needs different signals than BNB.
60 Bayesian optimization trials per token (max_depth, learning_rate, subsample, gamma, regularization). Objective function is out-of-fold logloss on the purged validation set — optimized for confidence calibration, not accuracy inflation.
A second XGBoost model (the “meta model”) is trained on out-of-fold predictions of the primary model. Its job: given the primary model says BUY — is this a good BUY? Output: a meta-confidence score (0–1). A signal only reaches the user if meta-confidence exceeds a per-token, per-side threshold tuned to achieve ≥62% holdout precision. Below threshold → HOLD, regardless of primary model strength.
A secondary optimizer finds per-regime thresholds on a rolling 600-bar out-of-sample window. Regimes: 3 volume tiers × 3 volatility tiers × 3 trend tiers = 27 buckets. The confidence bar for a BUY in “high volatility, downtrend” differs from “low volatility, uptrend” — automatic adaptation, no manual tuning.
Even after meta gate clearance, 8 structural filters run: (1) AI probability & class gap, (2) S/R proximity gate, (3) Reversal pattern scoring, (4) 3-candle confirmation, (5) Confluence vote, (6) Positive expected value required, (7) Macro regime filter, (8) BTC correlation brake.
A dual-layer user-protection system added above the model gates. First, an edge-score quality floor (0–100 scale, minimum 55) blocks low-confidence signals that the predictor threshold alone lets through — catching scale mismatches and low-conviction misfires before they reach the virtual wallet or user. Second, a beginner onboarding layer: pre-trade checklist popup on every session open, Guardian playbook with paper-trading-first philosophy, 1–2% risk rule, SL discipline, and a full in-dashboard Risk & Capital Management guide with capital tiers, position sizing formula, and R:R table.
Engine scans all 63 tokens on 1h candles, continuously, on Railway cloud. Live prices stream via Binance’s WebSocket mini-ticker (~1s updates), pushed to dashboard via FastAPI WebSocket. Signals written to Firestore for persistent storage and real-time sync. Frontend receives ticker updates every 100ms, full signal payloads every 500ms. Virtual paper wallet tracks open positions with TP/SL exit logic and syncs to the public track record on every close. Stack: Python 3.12, FastAPI, XGBoost, ccxt, Firebase, Railway.
AEGIS operates on a recurring subscription model with a 3-day free trial (no credit card) as the primary acquisition funnel. Revenue scales directly with subscriber count — near-zero marginal cost per additional user.
| Feature | Trial (3 days) | Basic · $3.60/mo | Intermediate · $7.20/mo | Pro · $14/mo |
|---|---|---|---|---|
| Tokens monitored | 63 (monitor) | 42 live | 58 live | 63 (all) |
| S/R levels & price targets | ✓ | ✓ | ✓ | ✓ |
| Plain-English analysis | ✓ | ✓ | ✓ | ✓ |
| AI conviction label | — | — | HIGH / MED / LOW | ✓ |
| Raw AI confidence % | — | — | — | ✓ |
| Fire signal + direction | — | — | — | ✓ |
| Guardian + Risk Guide | ✓ | ✓ | ✓ | ✓ |
| Multi-timeframe access | 1h only | 15m, 30m, 1h | All TFs | All TFs |
We believe the most trustworthy presentation is explicit about current limitations. Each weakness has a concrete mitigation already in development or shipped.
Early versions capped at 6 concurrent virtual positions with a 48-hour hold, limiting closed trade throughput to ~3 per day.
Done: Portfolio guard expanded to 12 concurrent positions (3 per correlation cluster), max hold reduced to 24h, and per-trade position size reduced to 7% base (from 10%) so more positions fit within the 60% capital cap. Expected throughput: 8–12 closed trades per day.
Meta-threshold values in meta files (0.6) were calibrated on the old 0–1 confidence scale, but edge scores are now computed on a 0–100 scale — meaning the predictor's fire gate was effectively never blocking anything.
Done: A hard edge-score quality gate (minimum 55/100) is now enforced at the live engine level before any virtual position opens. Low-confidence misfires (e.g. edge=3.0) are blocked regardless of the predictor's fire=True output.
The public live track record is growing but not yet statistically significant. Backtesting is rigorous; live performance takes time to compound at scale.
Fix: Transparent, unedited public track record accumulates automatically at 8–12 closed trades per day. All trades (wins and losses) are shown. Statistical confidence will compound over the coming months.
AEGIS identifies high-probability setups. It does not place orders. Execution, position sizing, and stop discipline are the user’s responsibility — introducing delay and emotional interference. New users are guided to practice on TradingView first.
Fix (Q4 2026): Exchange API integration for optional one-click execution via Binance and Bybit. Users retain full control; auto-execution is opt-in.
XGBoost models are trained on historical data. In genuinely unprecedented conditions (rapid regulatory change, black swan events), a model’s patterns may degrade before the next scheduled retrain.
Fix (Q3 2026): Automated drift detection. If live precision diverges from backtested expectation, a retrain triggers automatically within 24 hours.
Current payment processing (Razorpay) is optimized for Indian users. International users face friction: no native USD/EUR billing, limited card acceptance outside India.
Fix (Q4 2026): Stripe integration for USD/EUR/GBP billing. Global expansion begins with English-speaking crypto markets: UK, US, Southeast Asia, MENA.
All 60 tokens with live prices, RSI, trend regime, funding rate, plain-English analysis. 24 tradeable models. Real-time WebSocket. Public track record.
Universal XGBoost trained on 10 tokens, deployed on all 60. 25 strategies + 20 feature-engine indicators. 3 modes × 3 risk profiles. Virtual paper wallet with real trade history. Separate track record.
Monitors live precision vs backtested expectation. Triggers automatic retrain when drift exceeds threshold. Zero manual intervention required.
API connection to Binance and Bybit (opt-in). AEGIS places orders automatically when signals fire. Removes execution delay entirely.
USD/EUR/GBP billing via Stripe. English-speaking markets: UK, US, Southeast Asia, MENA. Affiliate / referral revenue share program.
Wallet flow data, exchange inflows/outflows, social sentiment as live model features. Portfolio risk dashboard: total exposure, cross-token correlation, combined max drawdown.
The retail crypto signal market is large, underserved, and dominated by low-quality products. AEGIS applies institutional quantitative methods — meta-labeling, per-asset models, regime-adaptive thresholds, rigorous backtesting — at a price point any retail trader can afford. The business is capital-light, recurring-revenue, and scales at near-zero marginal cost. The product is live, real, and the track record is public. Every trade. Every loss. No filters.