Cerebro: AI-Driven Crypto Portfolio Platform

Multi-agent portfolio intelligence for crypto: manipulation-resistant signals, tail-risk-aware allocation, and conversational research across 10,000+ tokens.
10,000+

Tokens tracked

< 60s

Research report generated

< 1s

Optimization solve time

7

Live data sources

About the project

Cerebro helps investors construct, analyze, and monitor crypto portfolios using machine learning and an LLM-powered research assistant, Alice.

KVY TECH built the platform end to end: real-time ingestion and entity resolution across seven market and social data sources, a manipulation-resistant momentum engine, a convex optimizer built around tail risk, and a multi-agent assistant that answers portfolio questions in natural language.

The ML recommender clusters 1,900+ tokens into 8 behavioural categories and tailors allocations to risk profiles from Ultra Conservative to Ultra Aggressive.

Client

Cerebro Pte Ltd

Engagement

Dedicated Team: AI/ML, Backend, Frontend, Data, DevOps

Service type

AI platform, conversational assistant & ML recommendation engine

Tech Stack

LangGraph, Python, Elixir, Node.js, Next.js, MongoDB, PostgreSQL, Qdrant, Redis, Modal, Vercel

The problem

Crypto investors, retail and semi-professional alike, face decisions across thousands of tokens with no reliable way to separate signal from noise. Existing tools lack personalisation, risk models suited to crypto, and real-time intelligence. The result is portfolios driven by emotion rather than data.

Information overload

Prices, social sentiment, breaking news, tokenomics and unlock schedules sit across dozens of fragmented sources. Synthesising them by hand is slow and error-prone.

Emotional decision-making

Without a data-driven framework, traders chase FOMO and panic-sell. A market that never closes amplifies the bias, producing poor entry and exit timing.

Risk models that do not fit crypto

Conventional tools assume normally distributed returns. Crypto returns are fat-tailed, so those models understate the downside precisely when it matters most.

Delayed signal detection

Breaking news moves crypto markets within minutes. Manual monitoring cannot match the speed required to act before prices adjust.

24 square layout grid 2

Market manipulation

Wash trading, pump-and-dump schemes and coordinated social campaigns distort standard momentum indicators, making manufactured activity look like genuine demand.

Engineering challenges

01

Signal quality

Many sources, low signal-to-noise. The platform needed robust features and a disciplined evaluation loop, not more data.

02

Manipulation resistance

Standard statistical measures are trivially gamed in crypto. The momentum engine had to stay reliable on adversarial data.

03

Explainability

Every recommendation needs a stated driver, constraint and trade-off, or users will not act on it.

04

Latency & scale

Sub-second optimization and sub-minute research over large historical datasets and continuous live feeds.

05

Security & privacy

Wallet addresses, balances and preferences protected end to end, including at the LLM boundary.

Our solution: a 5-layer architecture

KVY TECH built a modular, explainable portfolio engine and multi-agent LLM system, structured in five layers that carry data from raw ingestion through AI analysis to actionable delivery.

Layer 1: Data ingestion

Seven live sources are ingested and normalised into a single source of truth. Prices stream in real time over Binance’s WebSocket feed for listed tokens, while CoinGecko and CoinMarketCap are polled every five minutes to keep the long tail of 10,000+ tracked assets current. Entity resolution eliminates duplicate references to the same token, person or organisation across feeds, so every downstream signal is computed on clean data.

Source Data type Refresh
Binance (WebSocket) Live price stream for listed tokens Real-time
CoinGecko & CoinMarketCap Price quotes and metadata for remaining tokens Every 5 min
X (Twitter) API v2 Curated social posts Every 1 min
News RSS / APIs Long-form articles Every 10 min
LunarCrush Social sentiment, galaxy scores Every 1 hour
Hyperliquid Perpetual market data Every 1 hour
Token Unlocks Vesting schedules, tokenomics events Every 24 hours
  • Entity resolution: Token references, people, organisations and blockchain networks are normalised into a unified knowledge graph.
  • Semantic search: News articles and social posts are embedded as vectors in Qdrant and retrieved by similarity with confidence scoring.
  • Cross-referencing: Market events, entities and sentiment signals are linked automatically to build context around every asset.

Three specialised engines turn clean data into actionable signals.

A1 Momentum Score: manipulation-resistant

A proprietary momentum indicator, scored 1 to 100 and recalculated every 15 minutes, combining price momentum, volume activity and volatility burst. The design decision that matters: robust statistics (median and median absolute deviation) replace mean and standard deviation, so wash trading and artificially inflated volume cannot drag the signal. Manufactured activity stops looking like genuine demand.

CVX Portfolio Optimizer: built around tail risk

Allocation is solved by convex optimization (CVXPY) with the objective of minimising Expected Shortfall (CVaR) rather than variance. Constraints cover short-selling, concentration limits, sector diversification, turnover caps and configurable anchor assets. The optimizer runs every 4 hours and solves in under one second.

Breaking News Detector: three-model consensus

Three independent LLMs (Gemini, Grok and DeepSeek) score each piece of content, and at least two must agree before it is classified as breaking news. Social posts are held to a higher confidence bar than long-form articles. Alerts reach users on Telegram with a severity badge and source attribution.

Alice is a conversational research system built on LangGraph with a ReAct loop. Routing between agents follows explicit graph paths rather than probabilistic selection, so every research flow is reproducible and auditable, a requirement in financial applications.

Six specialised agents:

  • Portfolio Creation: Generates recommended token allocations from user inputs.
  • Category Adjustment: Rebalances sector weights such as L1s, DeFi and memes.
  • Token Discovery: Identifies outperforming or trending tokens.
  • Performance Simulation: Runs backtests and what-if scenarios.
  • Rebalancing: Suggests adjustments based on drift and risk.
  • Risk Evaluation: Analyses drawdown scenarios and correlation risk.

On receiving a research question, Alice decomposes it into three parallel angles (Market, News and Tokenomics), runs them concurrently, and synthesises a structured report with citations in under 60 seconds.

Questions Alice handles:

  • “What if I rebalance to 60% majors, 20% AI, 20% memes?”
  • “What is my portfolio’s correlation with BTC?”
  • “What is my drawdown in a 20% crash?”
  • “Can you shift my allocation to be more defensive?”
  • “Which tokens are outperforming right now?”
  • Admin dashboard (Next.js): Token explorer across 10,000+ assets, entity graph visualisation, news browser with semantic search.
  • Operational dashboards (Streamlit): Monitoring for data ingestion health, momentum scores and optimization results.
  • Telegram alerts: Real-time breaking news notifications with market context and severity.
  • Alice chat UI: Conversational interface for natural-language portfolio research.

Technologies used

Runtime

AI / ML

Databases

Frontend

Data processing

Deployment

Observability

High-level architecture​

Ingestion

Binance WebSocket for real-time prices; scheduled collectors for market, social and tokenomics feeds.

Agent

LangGraph ReAct loop with tools for data retrieval, optimization and alert rules.

Data

MongoDB for prices and analytics; PostgreSQL for relational entities; Qdrant for vector search.

Compute

Python on Modal for scheduled jobs; Node.js / Next.js API for the application; Redis for queues and caching.

Backtesting

bt library with BTC benchmark comparison and reporting.

Observability

LangSmith traces for agent conversations; centralised logs, metrics and errors.

Deployment

LangGraph Cloud for agents; Vercel / Northflank for web; containerised workers for schedulers.

Key technical decisions

Key technical decisions

1/7

Streaming every one of 10,000+ tokens in real time would be prohibitively expensive. Binance’s WebSocket feed covers listed tokens live, while the long tail is polled every five minutes, giving real-time precision where it changes decisions and efficient coverage everywhere else

Tiered price ingestion

2/7

Mean-variance optimization understates crypto risk because it assumes normally distributed returns. CVaR targets the tail events that decide outcomes in volatile markets.

Expected Shortfall over variance

3/7

Median absolute deviation replaces standard deviation, nullifying the effect of wash-trading outliers on momentum signals without sacrificing responsiveness.

Robust statistics for momentum

4/7

No single LLM is reliable enough for breaking-news classification. Requiring consensus across three independent models cuts false positives sharply while holding recall.

Multi-model voting for news

5/7

Explicit graph paths instead of probabilistic routing keep research flows reproducible and auditable.

Deterministic agent routing

6/7

Modal.com scales with scheduled jobs and API load, so the client pays for compute used rather than idle infrastructure.

Serverless compute

7/7

With seven external feeds and three LLMs in the loop, failures are certain. Every component is designed to return partial results rather than fail outright.

Graceful degradation

Key capabilities

24 briefcase 24
Risk-aware portfolios

VaR at 95% confidence, drawdown limits and cluster caps aligned to each user’s risk tier.

24 square layout grid 2

Editable suggestions

Users swap tokens or categories and the engine re-solves risk and return in under a second.

24 coins chart

Scenario Q&A

Trim to five tokens, shift defensive, test BTC correlation, model a 20% crash.

24 virtual space

Virtual portfolios

Create, save and compare what-if allocations with donut charts and KPIs.

24 server rack
Backtesting & metrics

Returns, volatility, Sharpe, Sortino and max drawdown against a BTC benchmark.

24 content delivery
Real-time alerts

Consensus-verified breaking news, threshold breaches and allocation drift, delivered to Telegram.

Security & privacy

Application‑layer encryption (AES‑GCM/Fernet/Tink) for sensitive fields (wallets, balances).

Prompt sanitization & masking for LLM I/O (LLM Guard) and prompt‑injection defenses (Rebuff).

Privacy‑first logging with hashing & redaction; scoped JWTs and row‑level security where applicable.

Runtime moderation and evaluation tests (Giskard) to prevent data leakage and regressions.

Results and impact

Market coverage

10,000+ tokens tracked across 7 live data sources with unified entity resolution

Price freshness

Real-time streaming for listed tokens, five-minute refresh across the long tail

Research speed

Structured, cited research reports in under 60 seconds, three angles run in parallel

Optimization speed

Portfolio allocation solved in under 1 second, re-run every 4 hours

Personalisation

8 behavioural clusters across 1,900+ tokens, matched to risk profiles from Ultra Conservative to Ultra Aggressive

Signal integrity

Momentum signals that hold up against wash trading and artificial volume

Alert precision

Breaking news verified by three-model consensus before it reaches a user

Explainable suggestions

Alice cites the drivers, constraints and trade-offs behind every portfolio change, so users can audit a recommendation instead of taking it on trust.

Production reliability

Cloud-hosted agents with LangSmith tracing give full visibility into agent behaviour in production.

Modular delivery

A layered architecture let the team ship and replace individual engines without destabilising the platform.

Need an explainable
AI portfolio engine?

We build reliable, auditable ML + LLM systems that users trust.