Naman Swami@kgfg0010095281
1
Events attended
1
Submissions made
India
2+ years of experience
About me
I am a creative vibe coder who focuses on high-level architecture and speed. I specialize in leveraging advanced AI agents and robust frameworks like FastAPI to rapidly stitch together high-performance applications. I believe in smart prompting, clean logic, and building bulletproof automated pipelines that get the job done efficiently.
🤓 Latest Submissions
ValueStream AI:Video Captioning Pipeline
Most modern video captioning pipelines suffer from meta-description leakage, cluttering social media captions with low-value, decorative noise such as describing a person wearing a hoodie sitting at a desk. ValueStream AI completely reimagines short-form video captioning by treating visual frames purely as informational filters rather than merely describing what a video looks like. By extracting high-density, actionable takeaways from combined audio transcripts and key visual assets, ValueStream AI synthesizes them into four distinct, high-converting social media tones: formal, sarcastic, humorous tech, and humorous non-tech. The project features an asymmetric multi-agent architecture running entirely on Fireworks AI serverless pay-per-token infrastructure, maximizing inference speed while maintaining an ultra-low cost per video run. In this pipeline, Agent 1 serves as the perception layer using MiniMax-M3 to interleave verbatim speech transcripts with structured visual metadata and enforce a deterministic category schema covering on-screen text, charts or diagrams, code snippets, product demos, and data overlays to eliminate background noise. Agent 2 handles creative synthesis via Gemma 4 31B IT, processing the noise-filtered document to synthesize four tone variants while dynamically optimizing outputs against rubric weights for hook strength and readability. Finally, Agent 3 acts as a dual-judge evaluator panel combining gpt-oss-120b as a text judge auditing factual grounding, schema compliance, and value density, alongside Qwen 3.7 Plus as a multimodal judge performing visual cross-examination across keyframes to verify visual claims and tone separation.
13 Jul 2026
