Software Design & Architecture
AVAL SYSTEMS
Hybrid AI financial reconciliation engine automating enterprise auditing.
OVERVIEW
A hybrid matching engine automating enterprise financial auditing by combining deterministic logic with semantic AI vectorization. Bridges structured accounting rules with unstructured transaction narratives to deliver near-zero manual reconciliation overhead.
KEY HIGHLIGHTS
- 01Embedded 768-dim transaction vectors via local Ollama LLM matched with pgvector cosine similarity
- 0212-use-case pipeline: Excel ingestion → PDF parsing → schema standardization → human-in-the-loop approval
- 03GoF patterns (Strategy, Template Method, Repository/DAO, Adapter) across fully Dockerized infrastructure
TECHNICAL IMPLEMENTATION
Semantic AI Integration
Embedded transaction narratives as 768-dimensional vectors via a locally-run Ollama LLM (nomic-embed-text) and matched them using Cosine Similarity search in PostgreSQL via pgvector. This allowed semantic matching of transactions that share intent but differ in wording, a problem deterministic rule engines fundamentally cannot solve.
Data Pipeline Architecture
Architected a 12-use-case pipeline covering Excel ledger ingestion (Apache POI), PDF bank statement parsing (Apache PDFBox), schema standardization, probabilistic match hypothesis generation, and a human-in-the-loop approval workflow. Each stage was isolated behind interfaces to allow independent testing and substitution.
System Design & Deployment
Applied GRASP and GoF design patterns — Strategy, Template Method, Repository/DAO, Adapter — across a strict multi-layer architecture. Containerized the full infrastructure, including PostgreSQL and the Ollama inference server, using Docker Compose for reproducible one-command deployment.
TECH STACK