Aman Nindra

Designing and building a self-hosted AI training platform to enable users worldwide to upload datasets (S3-style), configure hyperparameters, and train models through a web interface. Implementing GPU-aware job scheduling via Kubernetes to dynamically allocate workloads across multiple Linux machines based on real-time VRAM and compute availability. Deploying a Raspberry Pi as a lightweight control node to host the site, manage job queues, and persist datasets to attached external storage.

Developing an autonomous electric bicycle capable of navigating real-world terrain without human control. Integrating ROS 2 on a Jetson Nano Super (edge AI GPU) for perception, localization, and control.

Real-time voice simulator for emergency dispatch training powered by the ElevenLabs Conversational AI API over low-latency WebRTC. Dynamic evaluation engine parses live STT transcripts to score dispatchers across 5 critical criteria and compute a final performance grade. Integrated @vis.gl/react-google-maps for an interactive location-guessing system using Haversine distance scoring.

Fine-tuned BERT-base for 28-class multi-label emotion detection; achieved Macro F1: 0.64–0.70 and Jaccard (samples): 0.59–0.67. Built per-emotion threshold calibration (0.01–0.99 grid search per label), replacing a single global cutoff and improving minority-emotion performance. Benchmarked against GoEmotions baselines (Macro F1 ~0.49) and achieved Macro F1 ~0.64, Precision ~0.52, and Recall ~0.84 in 15 epochs.

Achieved Top 4 ranking (log loss: 0.6695) using EVA-Large (300M parameters) Vision Transformer on 16.5K camera trap images across 8 species. Engineered a production-grade training pipeline with DistributedDataParallel (DDP) across 4 Tesla V100 GPUs on AWS SageMaker with mixed-precision training, gradient clipping, and class-weighted loss. Optimized generalization to a -0.64% train-val gap (91.27% val accuracy); outperformed baseline ResNet18 by 3.1%.

A–Z ASL recognition with emotion-aware voice chat; ONNX Runtime enables client-side inference at 90 FPS. Trained CNN in PyTorch with augmentation and mixed precision; exported to ONNX with dynamic axes. Built Next.js UI (App Router, Tailwind) with accessible components and graceful fallbacks for camera and mic permissions.

Containerized full-stack app that ingests PDFs and returns resume recommendations, job matches, and a personalized dashboard to manage saved listings. Engineered a Firebase Realtime Database caching layer for job queries, reducing redundant Apify API calls and improving frontend response times. Designed a daily background cron scheduler (APScheduler) with MD5 signature hashing to auto-refresh listings and prevent duplicate database entries.