Generative systems across image · video · audio · language
I build end-to-end AI pipelines — from research and fine-tuning to production deployment — across diffusion models, speech, retrieval, and computer vision.
DiffSinger pipeline for an Ogilvy × Cadbury campaign that reached 14,000+ users at Christmas 2024. Phoneme alignment, variance modeling, and acoustic synthesis on a 3-hour custom dataset, with real-time inference on AWS Lambda.
Deployed FaceFusion for internal and marketing use. Accepts a per-segment JSON faceswap config, applies it frame-accurately, and stitches output with ffmpeg.
Natural-language search over video via open-vocabulary detection. Frame sampling + a YOLO-World inference pipeline for dynamic classes with no retraining, exposed through upload/query/timestamp APIs.
Full-stack RAG over 11,000+ SIGGRAPH paper chunks. Hybrid retrieval (semantic + BM25), Cohere re-ranking, cited LLM answers, and real-time SSE streaming.
Local-first multimodal video search for editors. Drop in footage, search in natural language, export timestamped clips — fully on-device. Dual retrieval (BGE transcript + OpenCLIP keyframe vectors) fused per video, evidence-gated chat that rejects uncited answers, and a 30-query eval harness that locks clip suggestions until top-3 accuracy ≥ 0.70.
Enterprise RAG for querying internal knowledge bases in natural language. Hybrid retrieval (embeddings + BM25), re-ranking, and RAGAS-based quality evaluation.
Led R&D on an English-to-Indian-regional-language speech POC. Benchmarked pipelines on latency, speed, and accuracy in a technical report; shipped on OpenAI's Realtime API.
End-to-end PDF parser for scanned and digital documents, handling mixed content — text, tables, and images — with PyMuPDF. OCR and layout extraction via Reducto and Qwen3-VL (through OpenRouter), normalized into a schema and saved as structured JSON for downstream LLM analysis. Async Celery + Redis pipeline for parallel processing with retry logic and API rate limiting.
Open-source Python package for scraping Pinterest images — search, download, HTML gallery generation, rate-limit handling, and a CLI. Published on PyPI.
Brave New Art (JioHotstar × Lenovo) — India's first fully AI-generated short film, produced end-to-end. Plus a Mother's Day film for SMFG India Credit.
Generated images in varied artistic styles. Fine-tuned FLUX, SDXL, and SD 3.5 LoRAs on Kohya_SS and published models to CivitAI.
Fine-tuned and published LoRA models on CivitAI across film, portrait, and product styles.
Agentic study companion for VideoDB's "Unlock the Footage" Global Media Intelligence Hackathon. Turns any lecture video into a full study pack — grounded clips, an AI-generated explainer video (FLUX visuals + TTS narration + auto-captions), flashcards, and summaries — from one natural-language prompt. A GPT-4o tool-loop agent picks the tools and streams live progress; optional Telegram delivery.
Desktop learning agent for educational videos (VideoDB Global Online Hackathon). A floating Electron overlay captures screen + audio via VideoDB, indexes streams with multimodal models, and writes structured notes into Notion.
A practical breakdown of identifying who spoke when in an audio stream.
Papers I read, break down, and implement.
Attention Is All You Need — 31 May '26
B.Tech in Computer Engineering · CGPA 9.22
8-week intensive by Dr. Arjun Jain (Jan–Mar 2026). Certificate ↗
Enhancing organisational security using augmented-reality smart glasses · App. No. 202321037263
Student Research Convention (West Zone) for ThirdEye — AR smart glasses for security.