01The problem
TryNectar provides companion, emotional-connection, and empathy experiences to general users and AI content creators across text, image, and video. Users expect a fast, coherent experience, but multimodal pipelines are naturally slow and inconsistent: appearance and tone can drift between generations, while every added modality increases latency.
The brief was to solve both problems at production scale without exposing the complexity underneath to users.
02Approach
I redesigned the generation pipeline around ComfyUI for image and video and LangGraph for orchestration, then addressed latency and consistency as separate engineering problems:
Latency
Optimized model inference on RunPod and restructured the pipeline so independent modalities run concurrently instead of waiting in sequence.
Character consistency
Used dynamic persona injection and context-window management so identity and tone remain stable across long conversations and multiple generated assets.
I also drove integration across model inference, content moderation, storage, and personalization into one platform architecture, while leading and mentoring the AI R&D team on engineering standards.