Foretale
No-Code Crypto Trading & Real-Time NLP Platform
Streaming architecture for market signals, sentiment, OCR, and visual automation
100K+
Daily Data Points
Processed across 7+ sources
<100ms
p99 NLP Latency
After isolating inference in Ray workers
25+
Automation Nodes
Custom Node-RED nodes for workflows
24/7
Continuous Pipeline
Streaming ingestion ran continuously
The Problem
Crypto trading strategies depend on social media, news, market data, whale movements, and image-based signals, but those inputs are fragmented across platforms. Most approaches require coding expertise and custom infrastructure for each strategy. Foretale set out to make those signals usable through visual workflow automation instead of requiring every user to build their own streaming system.
The Solution
Built a production streaming architecture with Kafka at the core, ingesting data from 7+ sources 24/7. The distributed ML backend (FastAPI/Ray/PyTorch) performs real-time sentiment analysis, emotion detection, and OCR extraction from images. A custom Node-RED fork provides visual workflow automation, allowing non-technical users to create complex data processing pipelines through drag-and-drop. Multi-tenant isolation used per-tenant secret handling with HashiCorp Vault. Lesson learned: Initially deployed NLP models directly in the main API process, which caused latency spikes during high-volume periods. Moving to Ray distributed workers with dedicated GPU allocation solved the contention issues and brought p99 latency from 800ms to under 100ms for the measured inference path.
Tech Stack
Backend
AI/ML
Frontend
Infrastructure
Data Processing
My Role: Co-founder & Lead Developer
- Co-founded and architected the complete streaming platform from data ingestion to action execution
- Built Kafka-based sensor network processing multi-source data streams 24/7
- Developed distributed ML backend (FastAPI/Ray/PyTorch) for real-time NLP inference
- Created custom Node-RED fork with 25+ automation nodes and Svelte-based UI components
- Implemented OCR pipeline for extracting text and signals from images at scale
- Built multi-tenant orchestration layer with Docker Swarm and HashiCorp Vault
- Designed Monte Carlo simulation engine for strategy validation
Platform Components
Streaming Pipeline
Kafka-based real-time data pipeline ingesting signals from social media, news APIs, and market data sources 24/7 with sub-second latency.
- Real-time Twitter/X stream processing
- Multi-source data aggregation (7+ sources)
- Event-driven architecture with idempotent processing patterns
- Horizontal scaling for burst traffic
NLP Inference Engine
Distributed ML backend performing real-time sentiment analysis, emotion detection, and text extraction from images at scale.
- roBERTa-based sentiment/emotion/irony classification
- EasyOCR for image text extraction
- Named entity recognition (NER) for signal detection
- Model-worker isolation reduced p99 NLP latency to under 100ms for the measured inference path
FlowStudio
Visual workflow builder enabling non-technical users to create complex data processing and automation pipelines through drag-and-drop.
- 25+ custom automation nodes
- Real-time flow execution with live data
- Svelte-based custom UI components
- Built-in simulation and validation
Key Differentiators
Streaming Architecture: Kafka-based pipeline handling 100K+ daily data points from 7+ sources
Real-Time NLP Inference: Distributed sentiment, emotion, irony, NER, and OCR workloads with <100ms p99 latency after worker isolation
Visual Workflow Automation: No-code interface for complex data processing pipelines
Multi-Source Data Fusion: Unified ingestion from social media, news, and market data APIs
Tenant Secret Isolation: HashiCorp Vault integration for per-tenant credentials
Want to discuss this experience?
I am open to full-time Senior AI/ML Platform Engineer roles where this kind of production AI, data, and platform work is useful.
Other Projects
Enterprise IoT Platform
Multi-Tenant Workflow Automation Infrastructure
Reduced customer onboarding time by 70% and eliminated 40+ hours/month of manual DevOps while the architecture was designed for 1000+ isolated workflow instances.
Zeitgaist
Cross-Lingual Social Intelligence
Synthesizes current insights from 6 platforms in seconds vs hours of manual monitoring, with cross-lingual search and filterable source context.