Moses Lua

Production trading intelligence platform processing 500+ data feeds (Polygon.io, Alpha Vantage, Finnhub, Reddit/WSB, X, TikTok, NewsAPI) normalized via Kafka with Rust/C++ hot-paths on TimescaleDB targeting sub-400ms latency.

• Dual LLM pipeline (Claude + FinBERT) for sentiment extraction and sarcasm detection
• 1,536-dim vector embeddings in pgvector using MEV-inspired similarity search
• Unicorn Scorer: 0–100 breakout signals via cosine similarity against 50+ historical templates
• Enterprise ML licensing to hedge funds; pre-seed raising $1.5–2M

Automated trading system for Solana DeFi, streaming mints via gRPC/Yellowstone, parsing Raydium/pump.fun layouts on-chain, executing sub-slot entries via Jito bundles with configurable bankroll management.

Multi-agent autonomous system for pricing and executing prediction market positions. Risk layer grounded in probability & statistics coursework. Each agent specializes in a stage with feedback loops for continuous improvement.

Proprietary multi-source data pipeline that identifies memecoin breakout probability by quantifying the variables that drive price action. Rather than relying on lagging technical indicators, delivers full analysis inline the moment you open a coin.

• Real-time social signal aggregation across multiple platforms
• On-chain flow analysis for whale wallet tracking and positioning
• Custom scoring engine based on actual breakout drivers
• Automated inline analysis delivery on coin pages

Using Markov chains and ML to model how snow levels affect energy markets. Pulling data from weather APIs, satellite imagery, road blocks, and airplane delays to predict moves in natural gas, power futures, and heating oil.

• Energy futures: NG (NYMEX), TTF, LNG, power (PJM, ISO-NE)
• Commodities: WTI, Brent, ULSD, heating oil
• Prediction markets on weather events
• Satellite and weather API integration

Building custom LLM training infrastructure using Unsloth for parameter-efficient fine-tuning at scale. Applied to financial sentiment analysis, autonomous trading agents, and multi-agent research coordination systems.

Architecture

• Base Model Adaptation: LoRA/QLoRA fine-tuning on Llama 3.1, Mistral, and Qwen2 architectures with 4-bit quantization for cost efficiency
• Domain Specialization: Custom training pipelines for financial text (earnings calls, SEC filings, macro news) and code generation (trading strategies, backtesting frameworks)
• Multi-Agent Training: Training coordinated agent systems where each agent specializes in research, execution, risk, or analysis stages
• RLHF Alignment: Preference learning with DPO/PPO for trading signal quality and research reasoning fidelity
• Curriculum Learning: Progressive training schedules that build complexity (syntax → semantics → reasoning → execution)

Training Infrastructure

• Multi-node distributed training with FSDP on custom GPU clusters
• Custom data pipelines: filtering, deduplication, quality scoring, synthetic data generation
• Evaluation frameworks: MMLU-pro, HumanEval, financial reasoning benchmarks
• Model serving with vLLM for sub-100ms inference at production scale

Contributed to AutoResearchClaw, an open-source autonomous research framework that achieved 10,000+ GitHub stars. The system autonomously produces academic papers through multi-stage pipelines with verified experiments, multi-agent peer review, and LaTeX export for major ML venues.

Pipeline Architecture (23 Stages)

• Phase A (Research Scoping): Topic definition, problem decomposition, research question formulation
• Phase B (Literature Discovery): Multi-source paper collection (OpenAlex, Semantic Scholar, arXiv), quality screening with GATE stages
• Phase C (Knowledge Synthesis): Topic clustering, hypothesis generation via structured reasoning chains
• Phase D (Experiment Design): Experimental protocols, code generation, resource planning with hardware-aware sandbox selection
• Phase E (Execution): GPU/MPS/CPU auto-detected execution, iterative refinement loops, result validation
• Phase F (Analysis): Statistical analysis, proceed/pivot/iterate decision making
• Phase G (Paper Writing): Full academic paper with LaTeX, multi-agent peer review (4-round audit)
• Phase H (Finalization): Quality gates, anti-fabrication verification, conference-ready export

Key Technical Features

• Multi-Agent Orchestration: Specialized agents for code generation, benchmark selection, figure creation, and paper writing coordinated via orchestrator
• Human-in-the-Loop (HITL): Six intervention modes from fully autonomous to step-by-step co-pilot with dynamic confidence-based pausing
• Sandbox Execution: Docker-based isolated environments with network policies, AST validation, and anti-hallucination claim verification
• Vector Search Integration: RAG over research corpus with cosine similarity scoring for literature relevance
• Cross-Domain Support: Pre-built adapters for math, CS, physics, biology, chemistry, economics with domain-specific skill loading

Showcase Results

• 8 published papers across 8 domains: random matrix theory, weak IV estimation, SIR/SEIR identifiability, Krylov preconditioners, LoRA optimization, token merging, lace exploration, and RL distillation
• 2,699+ tests passing with continuous integration
• +18.3% robustness improvement via MetaClaw cross-run learning integration