Aditya Arya

LFDT Hiero (formerly Hedera)

• –Contributing across the Hiero C++, Python, and Go SDKs since October 2025 — 48 merged pull requests spanning transaction APIs, JSON-RPC/TCK conformance endpoints, client abstractions, and CI tooling.
• –Promoted to Committer, the project's trusted-review tier — now reviewing incoming pull requests and onboarding new contributors alongside continued authoring work.

NextFlyt Aerospace

• –Built a semantic segmentation pipeline (ResNet50-UNet) on aerial drone imagery for terrain classification, reaching 85.6% pixel accuracy across 24 classes; diagnosed a spatial-downsampling bug causing small-object classes to vanish before reaching the encoder.
• –Compressed the segmentation model via FP16 quantization, cutting model size 66.5% (373MB→125MB) and inference time 14% (765ms→661ms) with 99.91% prediction agreement to the FP32 baseline.
• –Profiled a ViT-based (DINOv2 backbone) 3D reconstruction model on 4GB VRAM hardware; identified the hard memory ceiling causing throughput collapse and authored a feasibility report with INT8 quantization and cloud-based distillation as viable paths forward.

AmberFlux EdgeAI Private Limited

• –Improved retrieval relevance by 20–25% by redesigning a RAG pipeline’s ranking stage with query classification and hybrid scoring (Reciprocal Rank Fusion + normalized vector/BM25 combination).
• –Built an evaluation suite on RAGAS Faithfulness and Context Precision metrics to catch retrieval regressions before deployment.

ToeHo AI Private Limited

• –Fine-tuned Gemma-3-1B-IT via Unsloth for a bilingual (English/Hindi) assistant; implemented memory tracking and state management to resolve multi-turn context loss.
• –Engineered automated scripts using Playwright and Selenium to scrape and curate large-scale social media content (from X, Insta, etc.) for model training. Applied machine learning techniques to analyze YouTube data (transcripts, comments) to extract key audience engagement patterns and content preferences.