Ingest & normalize
Transaction payloads enter via Go/Gin APIs. Fields are normalized into a consistent risk schema so rules and models never depend on client-specific shapes.
Case study 01 · Fraud detection
Asguard
Real-time transaction risk scoring with rule engines, AI assessment, and investigator-ready natural language summaries.
Role
Fullstack / AI engineer
Stack
Go · Gin · Firebase · LLM
Domain
Fintech risk
The problem
High-volume payment and transfer flows cannot wait on batch fraud reviews. Investigators also cannot act on raw model scores alone — they need structured severity, category labels, and a plain-language explanation of why an event looks suspicious.
The engineering constraint is dual: keep decision latency low enough for live transaction paths, while preserving enough signal depth that false negatives don’t silently pass fraud and false positives don’t drown ops teams.
Design target
Sub-second risk evaluation with explainable outputs
The approach
Asguard uses a layered pipeline: deterministic rules catch known high-risk patterns quickly; AI risk assessment enriches ambiguous cases; and an LLM layer produces natural-language summaries for flagged events so humans can triage without replaying every field.
Events are auto-categorized by severity. Clear-cut decisions short-circuit early; borderline traffic pays the cost of deeper evaluation. Persistence and realtime surfaces (Firebase / relational storage) keep operational state queryable after the hot path returns.
Architecture
Stage 01
Stage 02
Stage 03
Pipeline stages and system responsibilities
Score & classify
Rule-based scoring runs first for known fraud vectors. AI risk assessment adjusts confidence for behavioral and contextual signals; results map to severity buckets.
Explain & persist
Flagged events receive LLM-generated summaries for human review. Outcomes and metadata land in durable storage / realtime channels for dashboards and follow-up.
Core engines
- ·Predictive / pattern layer — historical and synthetic patterns feed risk vectors before pure reactive review.
- ·Behavioral signals — anomalies relative to expected user or session baselines elevate score.
- ·Narrative layer — LLM summaries turn numeric risk into investigator-facing prose.
Technical foundry
Services
Go, Gin, Node.js (I/O paths)
Inference
Groq / Ollama-class LLM APIs
Data
PostgreSQL · Firebase
Infra
Cloud deployment (e.g. OCI)
What success looks like under load
40%
Reduction in successful fraud incidents (target metric)
<1s
Decision path latency budget for live scoring
99.99%
Availability goal for the risk service surface
Engineering takeaways
- · Separate hot-path rules from expensive model calls so most traffic stays cheap.
- · Treat explainability as a first-class product feature for fraud ops, not a log afterthought.
- · Persist severity + narrative together so dashboards and SLAs share one source of truth.
Azeru
Enterprise RAG & document intelligence