04 — Selected work

Models & systems, shipped.

RAG assistants that turn chats into leads, an AI editor that turns raw footage into YouTube Shorts, open-source image editing, autonomous agents and vision models — each measured by what it does in production.

P/01RAG · Vector search · LLM · Lead capture
Chat → LeadAnswers questions, captures prospects

RAG Lead Assistant — AI chat that answers & captures leads

Retrieval-augmented chat assistant grounded in business knowledge. It answers visitor questions accurately from real documents and captures qualified leads — name, contact and intent — naturally inside the conversation.

  • RAG Systems
  • LLM
  • Vector DB
  • Lead Generation
  • Chatbot
P/02Speech AI · Multimodal understanding · Auto-render
Raw → ShortCuts, captions & highlights — automated

AI Shorts Editor — Raw footage to YouTube Shorts

Upload raw clips and a script; the AI understands the video and audio, transcribes the speech, chooses the cuts, and renders vertical YouTube Shorts with animated captions and highlighted keywords.

  • Video AI
  • Speech-to-Text
  • Auto Captions
  • Multimodal AI
  • YouTube Shorts
P/03Diffusion · Segmentation · Inpainting
Open-sourceSelf-hosted generative image editing

AI Image Editor — Prompt-driven editing on open-source models

Carvona — an image editing pipeline built entirely on open-source diffusion and segmentation models. Select, remove, replace and restyle parts of an image with plain-language prompts, including professional automotive photo editing.

  • Stable Diffusion
  • Inpainting
  • Segmentation
  • Open-Source Models
P/04LangGraph · LangChain · Cognitive loops
1.2sMulti-step planning

Nami Agent — Autonomous multi-agent loop

Autonomous workflow engine with dynamic tool-calling, calendar orchestration, and self-correcting agent execution loops.

  • Agentic AI
  • LangGraph
  • LangChain
  • Python
P/05Qwen3 · FireRedOCR · Vision-Language
280msStructured JSON extraction

Buzz Card — Multimodal VLM document parser

Vision-Language pipeline extracting business cards and unstructured documents into CRM entities zero-shot.

  • VLM
  • Qwen3
  • FireRedOCR
  • Computer Vision
P/06Deep CNN · OpenCV · Biometrics
99.2%Anti-spoof accuracy

Face Anti-Spoofing — Real-time liveness detection

Production biometric security model distinguishing live human faces from 2D photos, masks, and screen replays in real time.

  • TensorFlow
  • CNN
  • OpenCV
  • Biometrics
P/07Stable Diffusion · ControlNet · YOLOv8
3.5sPhotorealistic inpainting

Diffusion Vehicle Studio — Segmentation-guided inpainting

Generative AI pipeline combining YOLOv8 segmentation with ControlNet to generate studio-grade vehicle visual presentations.

  • Stable Diffusion
  • ControlNet
  • YOLOv8
  • Inpainting
P/08FAISS · Dense embeddings · Gemini 1.5
CitedGrounded citation retrieval

Legal AI & RAG Engine — Grounded legal intelligence

Legal intelligence platform indexing Constitution & IPC legal documents with dense vector search for zero-hallucination answers.

  • RAG Systems
  • FAISS
  • LangChain
  • Vector DB
P/09Neural vocoder · Speaker conditioning
Zero-shotVoice prosody transfer

Neural Voice Cloning — Speech synthesis

Deep audio synthesis pipeline implementing speaker embeddings and neural vocoders for natural conversational speech.

  • Audio ML
  • TTS
  • Neural Voice
  • PyTorch
05 — Open source

AI repositories

github.com/prasannab4362
Contribution streakPrasanna B GitHub contribution streak statistics