Skip to main content
ShrirangMahajan
03Experience

A short résumé.

Four roles, four years, one obsession: shipping production ML systems that work outside the notebook.

  1. Machine Learning Engineer - II

    Skylark Labs
    • Sole ML engineer since September 2026, working with 3 SDEs and a product manager.
    • Video analytics platform (Kepler): built most of the CCTV and drone analytics (fire/smoke, PPE, ANPR, weapons, intrusion, crowd counting) and face recognition, served over gRPC on multiple hardware backends (Triton, Jetson, ONNX, CPU). Scales from 4 cameras on a Jetson AGX Orin to 40 on an A5000 cluster at 400 ms P99.
    • Inference performance: cut retrieval latency 88% (971 → 113 ms) by searching Qdrant embeddings only around detected objects and caching features to skip 87% of repeat model calls. Doubled throughput (3 → 7 FPS) with TensorRT on NVIDIA Triton, at 120+ inferences/s under 50 ms P95.
    • Aircraft tracking system: owned end to end as sole engineer, live at an operational airfield. Tracked 1,400+ takeoffs and landings with the aircraft in frame 90% of the time, at 22 FPS, using YOLO, visual odometry, SIFT/RANSAC pose correction and predictive PTZ control.
    • Edge deployment: designed a TensorRT gRPC inference server for road-damage detection (YOLO11) on Jetson AGX Orin, shipped as a .deb/systemd service. Cut per-frame time 2.7× (65 → 24 ms), mostly by rewriting preprocessing.
    • FrameSeek (prototype): natural-language CCTV search where any LLM (Claude, GPT, Gemini, Groq) queries an offline video index through 19 MCP tools. Index built with YOLO26, ByteTrack, SigLIP2 and DINOv2.
    • Drishti: built the team's annotation platform with GPU pre-labelling on Celery workers and automated QC, tripling annotation throughput (Django, React, PostgreSQL, Docker).

  2. Data Scientist

    Emergys
    • Led development of an LLM agent for BI analytics that turns plain-English questions into SQL and charts, cutting query turnaround 90% to under 5 s. Deployed to multiple enterprise clients as a core revenue product.
    • Built a promo-sales forecasting model (gradient boosting, R² 0.93) on Azure Databricks, with TPE hyperparameter search, to plan promotions.
    • Owned a price-outlier detection system end to end, from clustering and supplier-concentration metrics to the stakeholder dashboard.

  3. SDE Intern

    Atomic Loops
    • Engineered backend REST APIs using Django REST Framework for diverse products, facilitating data generation and collection for training ML models. Designed and implemented app for Yoga pose detection using Flutter.
    • Implemented a Real-time Chat API using WebRTC for a client's website.
    • Designed and implemented app for Yoga pose detection using Flutter.

  4. ML Intern

    Nishikawa Communications Pvt. Ltd.
    • Developed an image straightener tool using OpenCV and TesseractOCR. This system automatically corrects titled or flipped images, delivering results in a rapid response time of 1-2 seconds.
    • Built Japanese character classifier, utilizing transfer learning to recognize Hiragana and Katakana characters. The model was containerized using Docker for easy deployment.
    • Increased efficiency by writing custom Python scripts for data analysis after getting data from Google My Business API by using pandas. Updated the older script to match newer version of the API.
Publications

  1. Continual Field-Adaptive Models (CFAMs) for Post-Deployment Physical AI

    arXiv · Co-author (4th of 10)
    • Ran the continual-learning experiments and benchmarks and wrote that section of the paper: models learn new skills after deployment without catastrophic forgetting.
    • Backward transfer of −0.5 pp vs −11.4 pp for LoRA, meaning accuracy on earlier tasks barely drops while new ones are learned.