Lucas Gold

AI Invoice Extraction

• Purchasing and finance use the application in their regular invoice workflow. It replaces manual entry and saves approximately 40 hours a month.
• The LLM identifies invoice structure and pdfplumber extracts numeric values. This separates layout interpretation from the numeric checks used to validate the output.
• A second model pass and independent parsing re-check extracted values, with line amounts and quantities checked against document totals. Confidence is measured before and after the retry to decide whether the second pass resolved the disagreement or the field should be flagged. About 3% of line items reach human review, and corrections made there are added to the validation set, so later changes are tested against cases that were previously wrong.
• Developed and validated against approximately 80 vendor formats, with 130+ distinct layouts processed in production since, none of them needing a template of their own.
• The desktop client provides progress tracking, CSV export and processing on worker threads. Errors are reported per invoice so one failed document does not stop a batch.

Inventory Knowledge Base

• Keyword matching handles exact product codes and brand names. Vector similarity handles descriptive questions and paraphrases.
• PostgreSQL with pgvector stores documents and embeddings. Retrieval combines normalized keyword and vector scores using a weighted sum, tuned on logged staff questions and synthetic variants.
• Recall is measured on queries logged after tuning, so the evaluation set is held out by time rather than by sampling. On those queries the expected product ranked first for 84%, in the top five for 92%, and in the top ten for 96%.
• Retrieval covers vendor product guides, the live inventory catalogue and internal training and policy documents.
• Answers cite page ranges in vendor guides or clauses in store policies. Internal document dates are shown, and policy answers link to the full document.
• Staff can browse by category, ask follow-up questions and revisit previous conversations. Queries are scoped to the store where they work.
• Product comparisons use tables; procedures use numbered steps. Product answers include the stocking department from the inventory data.
• Question history supplies real queries for retrieval tuning, with synthetic variants adding alternative phrasings. Because the log keeps growing, each round of tuning is measured against questions asked after it — the product supplies its own held-out set.
• Responses include thumbs-up/down feedback. Procedures requiring a manager’s decision tell staff to escalate.

White-Label Restaurant Commerce Platform

• A shared React Native and React codebase serves iOS, Android, web, kiosk and POS. Branding is configured per restaurant; platform-specific files handle differences in kiosk and web screens.
• Branded apps and kiosks send orders to restaurant POS systems. Orders from external channels return to the platform dashboard for reporting and invoicing.
• Implemented Deliverect channel and POS APIs, including catalogue updates, order acknowledgements and availability controls, working with Deliverect’s team. The hub provides access to 40+ POS systems.
• Integrated Square directly using per-restaurant OAuth, catalogue synchronization, order APIs and Square Loyalty. Firestore and GCP Cloud Functions support real-time updates and webhook processing.
• Implemented Stripe Terminal and Square hardware payments alongside web and mobile checkout, including Apple Pay and Google Pay.
• Integrated DoorDash Drive with JWT authentication, store provisioning, delivery quotes at checkout and status webhooks for tracking and notifications.
• Used Stripe Connect to split payments when one order contained items from multiple restaurant brands.
• Built loyalty, rewards, gift cards and referrals, along with discounts supporting redemption limits and minimum spending requirements.
• Added receipt printing from kiosk and POS applications to Star and Epson hardware.

nextgenkitchens.comAndroid appiOS appNext Eats iOSWeb app