Mandeep Jaglan.

Video Processing & ML Infrastructure

• Built a RabbitMQ + Redis based cron pipeline ("fizzing") to orchestrate VLM inference on 2,000–4,000 videos per day, generating structured insights and result sheets.
• Reduced end-to-end pipeline runtime from ~1.5 days of manual processing to under 4 hours via fully automated parallel execution across 7–8 GPUs on 2–3 machines simultaneously.
• Developed GCS video management scripts for merging, cutting, and preprocessing timestamped video segments prior to inference.
• Built an end-to-end Jenkins CI/CD pipeline — from video recording through VLM inference to results delivery — with step-wise email notifications and queue polling.
• Engineered supporting Jenkins jobs: abort/delete stuck queue items, GCS upload/download, dynamic YML generation for Docker startup based on machine type and available GPUs, and data upload to an analytics dashboard.

Basketball Minimap Pipeline (Computer Vision)

• Designed and built a multi-stage CV pipeline transforming basketball broadcast video into top-down 2D minimap overlays using YOLOv9/v11, SAM 2, BoxMOT, and homography transformation — from unusable prototype to production-ready pilot.
• Implemented Global Track Association (GTA) to maintain consistent player identities across scenes and occlusions.
• Built team ReID using KMeansConstrained on HSV histograms; jersey OCR pipeline using FSRCNN super-resolution cross-validated against substitution logs.
• Applied bidirectional Kalman filtering on court keypoints and linear interpolation of player tracks (up to 45-frame gaps) for smooth minimap rendering.
• Produced analytics output: shot/turnover location maps, assist arrows, player marking analysis, Gaussian heatmaps, and per-player speed/distance tracking.
• Integrated a DeepLabel review loop for high-accuracy labeling; designed a SAM 2.1 targeted labeling system to reduce manual effort on future iterations.

NLP & Conversation Analytics

• Built an NLP → intent → BigQuery SQL pipeline for querying fizzing result sheets, with schema-aware prompt engineering, pre/post-processing, and automated chart and insight generation.

• Improved OCR accuracy to 95%+ across 10+ client datasets by optimising LLM configuration, prompt engineering, and building Python pre/post-processing scripts.
• Fine-tuned a GPT-3 based model for named entity recognition, increasing name identification accuracy by 35%.
• Automated data workflows using the Nanonets API and AWS S3 to handle 10 TB+ of data efficiently.
• Built an NLP-driven recruitment search tool (beanbag.ai) converting natural language queries to Elasticsearch and SQL via the Claude API, with PDL person API integration.

• Built a Facebook Ads data retrieval solution using APIs and web scraping, reducing processing time by 86%.
• Wrote scripts to collect and filter raw panel data from Hadoop, enabling access to 36+ TB of unstructured data.
• Automated data analysis and reporting with Python, SQL, and Snowflake; created and debugged 100+ SSRS reports.
• Developed a PySpark logistic regression model predicting panelist churn, improving prediction accuracy by 25%.
• Built an ML clustering solution to categorize users by web activity, informing UX and marketing strategies.