Hitesh Derkar
AI Engineer proficient in Python and JavaScript
New to Contra
Hitesh is building their profile!
Won the DataHack challenge by orchestrating a spatial-temporal action recognition pipeline utilizing OpenPose and a Multi-Domain Regularized Graph Convolutional Network (MDR-GCN) to analyze skeletal sequence chunks, accurately extracting and time-stamping complex figure skating choreography into structured JSON data.
Developed an automated multi-modal mapping algorithm that dynamically translated kinematic predictions into contextual, segmented text prompts, successfully matching specific physical movements (e.g., jumps, spins) with targeted musical orchestration parameters.
Integrated the InspireMusic generative AI stack utilizing text-to-music and temporal continuation modes, chaining dynamic segments to synthesize seamless, multi-minute 48 kHz stereo audio tracks perfectly synchronized with the input video.
0
41
Engineered a Sequence-to-Sequence (Seq2Seq) symbolic music generation pipeline utilizing U-Net, Bi-LSTM, and Encoder Transformers to synthesize simplified MIDI and sheet music, achieving an 89% F1-score in difficulty-level prediction.
Architected a scalable ETL pipeline to parse and process 440 complex temporal files from AWS S3, engineering a novel, time-aligned dataset specifically designed for difficulty-controlled Music Information Retrieval (MIR).
Developed temporal synchronization algorithms to align machine-transcribed audio stems with simplified symbolic sequences, establishing complex multi-modal mapping to train robust token-to-token translation models.
1
0
53
Engineered a custom AI workflow from the ground up, fine-tuning Llama 3.2 via LoRA to generate rule-compliant mechanics, and directly deployed the model into a real-time Unity environment. Executed rigorous version control and continuous testing workflows, balancing algorithmic optimization with
pragmatic deployment thresholds to meet production deadlines.
1
0
118
Multi-agent AI framework (Next.js, Python, SQLite) that orchestrates autonomous LLM personas in a strict peer-review simulation. By enforcing rigorous, stateful debates between agents, the system resolves epistemic drift and archives only consensus-driven, validated research findings into a persistent knowledge base.
0
125
Precision farming AI platform that processes drone imagery using a custom YOLOv8 model for weed detection. Integrated a multi-agent LLM architecture and live weather APIs to generate real-time agronomic insights.
1
0
141