Neptune Booking

Multi-tenant scheduling SaaS with an LLM onboarding agent, embedding-powered search, and an NLP booking parser at its core.

year

= 2026

role

= Designer, ML / Full-Stack Engineer, Founder

Neptune is a booking platform I designed and built end-to-end — but the part I'm proudest of is the ML layer woven through it. A GPT-4o-mini agent in JSON mode turns a paragraph of plain text into a fully working business config (hours, services, resources, pricing, branding). A semantic-search layer built on OpenAI text-embedding-3-small + MongoDB Atlas Vector Search lets customers describe what they want in their own words. An NLP booking parser turns 'patio table for 4 around 6:30 tonight' into structured slots. All of it runs on real bookings, every day.

• → LLM onboarding agent (gpt-4o-mini, json mode) — plain text → validated booking config in one call, with tab-aware follow-ups when key facts are missing
• → Semantic service search on text-embedding-3-small + MongoDB Atlas $vectorSearch, scoped per tenant; smart re-embed only when canonical service text changes
• → NLP booking parser (Microsoft Recognizers + custom keyword/regex matching) for dates, times, ranges, durations, party size, resource preferences, and AM/PM disambiguation against business hours
• → Conversational admin agent: multi-turn, tab-aware system prompt, returns JSON diffs with adds / edits / tombstone deletes — merged atomically by id
• → Eval harness across 8 business archetypes (salons, lanes, courts, spas) with structural + cross-reference validators and retry-on-truncate logic
• → End-to-end production system: Stripe payments, Google Calendar sync, embeddable widget, real-time admin dashboard, sliding-scale pricing engine