Nikunj Rajesh Lad

Computer Vision Developer | Machine Learning Engineer

© 2026. Credits.

Experience

Knightscope Inc.

Jan 2021 – Present
Sr. Machine Learning Engineer · Sunnyvale, CA
Sr. Machine Learning Engineer Mar 2025 – Present
  • Contributing to a natural-language search and question-answering system over security-incident imagery from the robot fleet: embedding every frame with a SigLIP dual-encoder and indexing in FAISS so an operator can retrieve incident images by plain-English description. Top-k retrieved frames are passed with the question to Qwen2.5-VL (served via Ollama) to generate grounded answers that cite the specific frames each claim is drawn from. Evaluation is designed around Recall@k / MRR for retrieval and a cross-family LLM-as-Judge (Claude API) for scoring correctness and grounding, with a human-validated calibration step before trusting it at scale.
  • Developed an auto-labeling system using multi-modal Vision-Language Models (OWLv2 / Qwen) to pre-annotate images for semi-supervised human labeling, reducing manual labeling effort by 50%.
  • Devised a DVC-based dataset versioning system with images in AWS S3 and labels in PostgreSQL RDS, tracking image/label version mappings in RDS to enable reproducible results and rollback-safe ML experimentation with MLflow.
  • Fine-tuned transformer detector RT-DETRv2-R50 and YOLOv7/v9/v12 models for person and vehicle detection on AWS EC2 g4dn.xlarge instances. These models run real-time inference on the Jetson ecosystem over 4x GMSL/IP camera input streams through a GStreamer pipeline leveraging accelerated NVIDIA DeepStream plugins, following NIST 800-53 security standards.
  • Improved ALPR (Automatic License Plate Recognition) accuracy by adding image pre-processing (Adaptive Gamma Correction, CLAHE, Laplacian and Sobel edge enhancement) to enhance and filter plates before OCR, plus a post-OCR consensus matching procedure using Levenshtein distance to cluster similar reads and Needleman-Wunsch alignment within each cluster to derive the consensus plate.
Machine Learning Engineer Jan 2021 – Feb 2025
  • Owned the AI/CV stack for real-time security-incident reporting on Jetson Xavier AGX and Jetson Orin NX modules across 4 product types, sustaining ~$5M in annual recurring revenue across 50+ clients nationwide.
  • Built the infrared thermal detection and streaming pipeline using the FLIR Lepton 3.5 camera, combining a YOLO-X object detector with Adaptive Relative Temperature Thresholding to deliver real-time localized temperature information, streamed via a GStreamer + RTSP server for user-facing thermal video.
  • Optimized real-time edge detection across the deployed fleet by quantizing object detection models to INT8/FP16 via TensorRT and adding a per-detector ByteTrack tracker to suppress duplicate alerts from spamming the server — reducing cellular costs by 20% and network bandwidth by 50%.
  • Packaged the inference pipelines as Dockerized services deployed via Docker Compose, with telemetry monitoring through Prometheus and Grafana.
  • Led live-server migrations from legacy AWS EC2 (Ubuntu 14.04) to g4dn.xlarge (Ubuntu 22.04), collaborating cross-functionally with the cloud team on a Blue-Green transition of servers and traffic for zero downtime, and building a FastAPI + Nginx server with self-healing workers serving ~1000 req/s from fleet robots.
  • Designed and implemented the consolidation of 3–4 disparate AI pipelines into a unified cross-platform inference stack compatible with Jetson Xavier AGX / TX2 (JetPack 4.6, Ubuntu 18.04) and Jetson Orin NX (JetPack 6.0, Ubuntu 22.04).

Squark

Oct 2020 - Dec 2020
Machine Learning Engineer
  • Worked on developing a system to observe and visualize Bias and ensure Fairness in machine learning systems using Statistical Parity Difference, Equal Opportunity Difference, Average Odds Difference, Disparate Impact and Theil Index Metrics.
  • Mitigating Bias and ensuring fairness using optimized pre-processing & adversarial debiasing algorithms.
  • Also working on reducing data dimension using t-Stochastic Neighbor Embedding (t-SNE) and developing interpretable models using Local Interpretable Model Agnostic Explanations (LIME).

UII America Inc.

Sept 2019 - Jan 2020
Computer Vision Research Intern
  • Implemented Dense Correspondence Estimation algorithm for real-time patient inference and increased inference speed by 50% to ~20-21 FPS using multi-threaded Dockerized architecture on Nvidia P5000 GPU
  • Developed a technique to acquire skeletal positions of human body by reading depth image frame from RGBD images captured using Astra Orbbec Depth Camera.
  • Annotated image data using Densepose for keypoint estimation and ran the system on NVidia P5000 GPUs.

Northeastern University

Jan 2019 - Apr 2020
Graduate Teaching and Research Assistant

InVideo

Sept 2017 - May 2018
Software Engineer
  • Scaled up video production rate for clients from 30 videos/day to 300 videos/day by implementing Geometric Perspective Transformation on image masks
  • Improved software efficiency by 60% by saving 10 mins of rendering time per video by incorporating Face Detection using Convolutional Neural Networks for automated text positioning
  • Parsed news articles for understanding language semantics using Spacy and NLTK for text-to-video translation
  • Integrated additional features for our product like providing users the ability to add GIFs, social media widgets, image icons, etc for contextual representation of information in videos.

K. J. Somaiya College of Engineering

May 2017 - Aug 2017
Research Assistant
  • Worked Under Dr. Jagannath Nirmal, HOD, Electronics Department, KJSCE on industrial Machine Learning and Computer Vision applications.
  • Worked on a project demanding the need to develop a system to classify Dysphonia patients from normal patients using Machine Learning. It uses an algorithm which uses a robust feature extraction technique named I-Vectors for classification. Entire implementation done in MATLAB and can be migrated into Python packages. Successfully modelled a system which gives 98% accuracy using the proposed technique as compared to the 92% before using the technique. Used Support Vector Machines (SVM) to perform Binary classification.1
  • Simultaneously worked on another project requiring the need to classify patients with Dysphonia, Paralysis, Laryngitis and Normal. Used the above proposed technique of I-Vectors to develop 3 different models to classify these 4 classes. Support Vector Machines (SVM), Naive Bayes and K-Nearest Neighbours (KNN) were used with multi-class functionality. The entire system was developed in MATLAB and can be migrated into Python packages.2