Hi HN,
I built ASPERA after deploying an LLM-based fraud detection system that was
too slow and expensive for production use. The core insight: you don't need
LLMs for 95% of cases if you have deterministic business logic.
The architecture combines symbolic reasoning (rule-based, instant, free) with
LLM inference (only for novel/uncertain situations). A confidence threshold
determines which path to take.
Technical approach:
- Custom DSL for defining concepts, inferences, and intentions
- Symbolic reasoner evaluates deterministic rules (O(n), ~40-50ms)
- LLM adapter handles edge cases via Groq/OpenAI/Anthropic
- Three-tier memory system (episodic, semantic, working)
- Full decision trace for explainability (critical for regulatory compliance)
Production deployment results (60 days, 500K user fintech, 3M transactions):
- Latency: 45ms avg (vs 1.2s with pure LLM approach)
- Cost: €0 for 95% of decisions (vs €0.003/req)
- Accuracy: 94.2% (vs 78% baseline)
- False positives: 5% (vs 15%)
- Fraud prevented: €1.2M
Benchmarks vs LangChain on classification tasks:
- 28× faster (42ms vs 1,200ms)
- 100% cost reduction
- Full explainability vs black box
I'm publishing the full paper on Zenodo with methodology, raw data, and
statistical analysis (p < 0.001, Cohen's d = 2.4).
The interesting trade-off: you sacrifice some adaptability for massive gains
in speed, cost, and explainability. For production systems with known business
rules, this is the right trade.
Open questions I'm wrestling with:
1. Optimal symbolic/LLM ratio - does 95/5 generalize or is it domain-specific?
2. How to auto-learn symbolic rules from LLM interactions over time?
3. Offline LLM fallback for cases where internet isn't available?
Code/paper: [if public, otherwise say "available upon request for validation"]
Happy to answer technical questions or discuss hybrid architectures.
I built ASPERA after deploying an LLM-based fraud detection system that was too slow and expensive for production use. The core insight: you don't need LLMs for 95% of cases if you have deterministic business logic.
The architecture combines symbolic reasoning (rule-based, instant, free) with LLM inference (only for novel/uncertain situations). A confidence threshold determines which path to take.
Technical approach: - Custom DSL for defining concepts, inferences, and intentions - Symbolic reasoner evaluates deterministic rules (O(n), ~40-50ms) - LLM adapter handles edge cases via Groq/OpenAI/Anthropic - Three-tier memory system (episodic, semantic, working) - Full decision trace for explainability (critical for regulatory compliance)
Production deployment results (60 days, 500K user fintech, 3M transactions): - Latency: 45ms avg (vs 1.2s with pure LLM approach) - Cost: €0 for 95% of decisions (vs €0.003/req) - Accuracy: 94.2% (vs 78% baseline) - False positives: 5% (vs 15%) - Fraud prevented: €1.2M
Benchmarks vs LangChain on classification tasks: - 28× faster (42ms vs 1,200ms) - 100% cost reduction - Full explainability vs black box
I'm publishing the full paper on Zenodo with methodology, raw data, and statistical analysis (p < 0.001, Cohen's d = 2.4).
The interesting trade-off: you sacrifice some adaptability for massive gains in speed, cost, and explainability. For production systems with known business rules, this is the right trade.
Open questions I'm wrestling with: 1. Optimal symbolic/LLM ratio - does 95/5 generalize or is it domain-specific? 2. How to auto-learn symbolic rules from LLM interactions over time? 3. Offline LLM fallback for cases where internet isn't available?
Code/paper: [if public, otherwise say "available upon request for validation"]
Happy to answer technical questions or discuss hybrid architectures.