mewatch AI - Conversational OTT Chatbot & EPG Engine
Searce Inc (Client: Mediacorp), 05/2026 – Present
Engineered a conversational OTT AI chatbot & hybrid search engine for Mediacorp's streaming platform (mewatch).
Architected a custom LangGraph language detector node with recency-prompt injection to handle 6+ languages (Tamil, Mandarin, Singlish, Malay, English, Chinese) with English fallback under cost-constrained Gemini 2.5 Flash models. Migrated EPG discovery from NL2SQL to Vertex AI Search with custom BigQuery data pipelines and engineered stateful recommendation filtering for VOD.
Impact
- Bypassed model-level attention decay in massive context windows by decoupling multilingual routing into a dedicated LangGraph node with dynamic recency-prompt injection, achieving 100% fallback accuracy on Gemini 2.5 Flash and reducing LLM token costs by 80%.
- Migrated EPG discovery from NL2SQL to Vertex AI Search over a 15-day rolling BigQuery dataset, supporting fuzzy title matches, temporal schedule filtering, and semantic queries while eliminating SQL syntax errors.
- Built stateful pagination and context-aware filtering for the VOD recommendation agent, achieving 0% duplicate show suggestions across continuous multi-turn sessions.
Stack
Python, LangGraph, Vertex AI (GCP), Cloud Run, Cloud SQL, BigQuery, Apigee, Model Armor, Cloud Trace