Production Systems Engineering at Tessact
Built production AI/ML systems serving JioHotstar, SunTV, and Jeevanvidya. Promoted to SDE II in January 2026 for architecting a face tracking service that reduced infrastructure costs by 40× and saved $2,500/month.
Tech Stack
Python
TypeScript
InsightFace
LangChain
LangGraph
Whisper
CLIP
Django
FastAPI
Next.js 14
React
Tailwind CSS
PostgreSQL
Redis
Docker
GCP Cloud Run
AWS Lambda
AWS MediaConvert
FFmpeg
RabbitMQ
Celery
Sentry
Face Analysis Service (Led to Promotion)
- Architected face analysis pipeline for long-form video: detection, identification, tracking, and clustering using InsightFace embeddings with ANN indexing.
- Replaced AWS Rekognition — cost dropped from $6 to $0.15 per video hour (40× cheaper) with 10× faster throughput.
- Generated $2,500/month savings at 100 hours/week scale, with headroom for 10× growth without proportional cost increase.
- Evaluated production ML models: InsightFace buffalo_l (SCRFD-10G-GNKPS detector, GLINTR100 embeddings), benchmarked clustering algorithms (HDBSCAN, Chinese Whispers, Agglomerative).
- Achieved 95%+ accuracy across film/TV content with stable identity tracking via Chinese Whispers clustering.
- Productized as reusable service on GCP Cloud Run with clean REST APIs — used by JioHotstar, Jeevanvidya, and SunTV.
Multimodal AI Content Automation Pipeline
- Built end-to-end AI pipeline reducing video editing time from 60 to 10 minutes (6× improvement).
- Designed multimodal signal processing: Whisper v3-turbo ASR, TransNet-V2 shot detection, CLIP visual tagging, and Tesseract OCR into unified data store.
- Implemented LLM agent system using LangChain + LangGraph to analyze combined signals and intelligently plan cuts and transitions.
- Scaled to bulk processing: supports batch operations on hundreds of videos simultaneously.
- Actively used by JioHotstar, Jeevanvidya, and SunTV for automated content curation.
Media Processing Infrastructure
- Architected queue-driven video transcoding: FFmpeg + RabbitMQ + Celery with NVIDIA L4 GPU acceleration (NVENC).
- Generated adaptive bitrate streaming: HLS/DASH at 1080p/720p with x264/x265 codecs.
- Implemented failover architecture: AWS MediaConvert backup for GPU pipeline failures, ensuring 99.9% reliability.
- Delivered successful PoCs helping win contracts with major media clients.
Frontend Rebuild
- Rebuilt product frontend from scratch using Next.js 14, TypeScript, and Tailwind CSS.
- Achieved 90% reduction in technical debt and 97% crash-free sessions (Sentry monitoring).
- Established best practices: TypeScript strict mode, comprehensive error handling, server components, optimistic updates.