AutoVaultLK

Many second hand vehicle dealers still enter listings manually slow, error-prone, and inconsistent across a large inventory. Existing local platforms only support traditional filter-based search, so buyers who don't know exact make/model/trim naming struggle to find what they want, and international platforms built for large businesses are too expensive for small dealerships.

The system lets dealers bulk upload inventory as CSV/JSON with a ZIP of images; an asynchronous ETL pipeline validates and normalizes the data with deterministic rules first, escalating only ambiguous rows to an LLM, and never publishing an AI inferred value without dealer review. Buyers search the resulting catalogue with structured filters, natural language, or both the query is parsed, embedded, and matched against listings using a combination of exact filters, semantic vector similarity, and fuzzy text matching, all resolved in one ranked SQL query.

Dealers register, get approved, and upload their inventory in bulk rather than entering vehicles one at a time. Each upload triggers an AWS Step Functions execution: the file is validated structurally, split into 100-row chunks, and each chunk is normalized against dictionaries and regex patterns for makes, models, fuel types, and numeric fields. Rows the rules can't confidently resolve an ambiguous make spelling, a price written as "around 9 mn," a fuel type only mentioned in free-text notes are batched and sent to Groq for LLM-assisted repair, with every returned field validated against a whitelist before it's trusted. Listings land in a PENDING_REVIEW state; dealers see exactly which fields were AI-inferred, with the model's reasoning, before approving them to go LIVE.

Buyer-facing search accepts structured filters, a free-text natural-language query, or both. A deterministic parser extracts hard constraints (seats, price, year) from the query text; anything left over is treated as descriptive "vibe" text and embedded with a MiniLM sentence-transformer for semantic ranking against pgvector. Typos in make/model names are resolved via PostgreSQL trigram matching against a dictionary table during parsing never against the live listings table keeping the ranking logic simple and the indexes usable. Filters gate the result set; only the vector score decides ordering within it.

The backend is five independently deployable services (Auth & User, Marketplace, Ingestion & ETL, Admin, Notification), all running as AWS Lambda functions behind API Gateway, sharing one PostgreSQL (RDS) instance with per-service schema ownership. The ETL pipeline alone spans eleven Lambda functions with deliberately uneven memory allocation 256 MB for lightweight validation steps up to 3 GB for the MiniLM embedding step coordinated declaratively through Step Functions rather than a single monolithic worker, so that an LLM outage degrades an upload to ~95% success instead of failing it outright. Infrastructure is defined in Terraform and deployed via GitHub Actions CI/CD, with local development running the full stack through Docker Compose.