Agharia Kasimali - AI & Machine Learning Portfolio

Welcome to my research portfolio! I am an enthusiastic and highly driven machine learning researcher, eager to solve complex, real-world AI challenges across diverse data modalities. My core interests lie at the intersection of deep learning, computer vision, and natural language processing.

This space serves as the primary interactive verification archive showcasing my cross-domain capabilities, encompassing both my remote sensing research at the Indian Space Research Organisation (ISRO) and my multilingual speech processing work with the Annam.ai & IIT Ropar Initiative.

Multilingual NLP & Speech Research

Automatic Speech Recognition (Annam.ai & IIT Ropar Initiative)

My research focuses on building scalable, equitable ASR models for 10+ low-resource Indic languages, handling complex acoustic characteristics across the Indo-Aryan and Dravidian language families.

View the Complete Multilingual ASR Archive ➔ Explore VoxLingua107 cross-lingual evaluations, Whisper fine-tuning logs for Malayalam, joint ASR+LID prototypes, and end-to-end IndicTrans2 inference pipelines directly in the browser.

Manuscript Under Review / Preprint

Dual-Track Rule-Based Bootstrapping & ML for District-Scale Crop Classification (ISRO)

Focus: Resolving multi-class label noise, dynamic K-Means phenological rule appeals, and operational XGBoost acreage reconciliation across multi-temporal Sentinel-2 imagery.

01: Pipeline Curation & XGBoost Training — End-to-end Track 2 rule engine calibration (Batches 1–3), Ground-Truth hard-case extraction, and final XGBoost model training logs.
02: Operational Deployment (Patiala District) — Track 1 district-scale inference, pixel-wise segmentations, and operational acreage reconciliation against official tabular targets.
03: Spatial Transferability (Fatehgarh District) — [Ongoing / Future Work] Zero-shot cross-district spatial transferability evaluation and agro-ecological boundary validation logs.

🛰️ Foundational Geospatial Vision Architectures & Prototyping

Interactive training runs, checkpoint matrices, and upstream spatial modules evaluated across multi-spectral satellite layers during my ISRO tenure:

1. Foundation Architecture Implementations (SatMAE Base)

SatMAE From Scratch (5000 Epochs) — Baseline autoencoder reconstruction logs.
SatMAE Optimized 10-Channel Framework — Fine-tuned multi-spectral tokenization configurations.
SatMAE Partial Vision Transformer Analysis — Structural optimization variants.
SatMAE Exact Code Execution Validation — Isolated single-batch consistency evaluations.
SatMAE Partial Checkpoints Matrix — Convergence behavior at 5k iterations.

2. Foundation Architecture Implementations (Prithvi Base)

Prithvi From Scratch (Final Run) — Complete weight updates and spatial adjustments.
Prithvi Core Foundation Pipeline — Basic infrastructure setup and layer definitions.
Prithvi Partial Epoch Evaluator — Tracking loss adjustments mid-cycle.

3. Early Prototyping & Upstream Operations

Punjab Rule-Based Classification Framework — Legacy rule scripts and target remote sensing masks.
Pixel-Wise Segmentations — High-resolution dense crop boundary prototyping.
Requirement Analysis Module — Upstream feature selection checks.

🔒 Security Protocols Maintained: To ensure institutional compliance, all cloud project tokens, GEE keys, and private infrastructure paths have been programmatically scrubbed and replaced with standard [REDACTED] tags.