Technical Research & Architectural Innovation

Bridging Scientific Discovery with Production-Scale Machine Learning

ASTRA: A Foundational Framework for Astronomical SSL

Self-Supervised Learning Transformer TensorFlow MirroredStrategy 8x NVIDIA A100 (80GB) NVIDIA Academic Grant

Astra-CLR: Architected the first multi-filter time-series Transformer for the ZTF survey. Pre-trained on ~2.1M unlabeled light curves utilizing TensorFlow MirroredStrategy across a cluster of 8x NVIDIA A100 80GB GPUs.

High-Throughput Inference: Successfully scaled the pipeline to generate representational embeddings for ~70 Million samples (N_obs ≥ 2000), establishing a foundational-scale dataset for international team-wide research.

The framework utilizes a novel multi-view late-fusion architecture and "multi-scale temporal view" mapping strategy (an adaptation from "multi-crop" paradigm) to learn robust local-to-global representations from inhomogeneous multi-filter light curves, achieving a 77% fine-tuned accuracy across 12 broader variability classes.

Ongoing Expansion: Spearheading the development of "non-contrastive" framework within ASTRA to establish a universal scientific foundation model for multi-survey photometric data and is slated for upcoming release.

🔬 View ASTRA Codebase 📄 Read Paper (arXiv)

Figure 1: The ASTRA-CLR framework utilizes asymmetric temporal views and a shared AstraNet encoder, consisting of a backbone F (·) and a projection head G (·), to maximize agreement in a shared latent space utilizing NT-Xent loss objective.

APPLIED AI & AGENTIC SYSTEMS

Sentinel-Hub (Multi-Agent System)

Orchestration • LangGraph • Multi-modal

Challenge: Enterprise security data is often siloed, making coordinated multi-modal attacks (e.g., a deepfake video paired with a server breach) difficult to investigate.
Solution: Architected a stateful, 4-agent cyclic MAS that orchestrates Vision, Log, and Financial specialist agents to cross-validate findings and generate unified forensic reports.

📂 View MAS Architecture

DeepShield (AI Safety & DPO)

DPO • Alignment • Computer Vision

Challenge: Baseline deepfake detectors suffer from high false-positive rates on authentic media and low sensitivity toward "seamless" manipulations.
Solution: Implemented Direct Preference Optimization (DPO) based PEFT approach to align a DINOv2 backbone with human-perceived risk, reducing False Positives by 51% while maintaining 97.5% sensitivity on high-fidelity fakes.

📂 View DPO Logic

LogSentinel (InfraSec)

Zilliz • BERT • Scaling

Challenge: Detecting "Zero-Day" infrastructure anomalies in massive, unreadable HDFS log streams.
Solution: Engineered a semantic guard using distilled BERT embeddings and Zilliz (Milvus) to identify out-of-distribution system behaviors in real-time.

📂 View Code

NeuralAudit (FinTech)

Supabase • SSL • Tabular

Challenge: Fraud detection in tabular data without historical labels.
Solution: Developed a Self-Supervised anomaly detector using bottleneck autoencoders to identify fraudulent transactions via Euclidean distance in high-dimensional latent space.

📂 View Code

CanFin RAG (Finance)

Temporal-RAG • Parsing • Markdown

Challenge: Standard RAG systems fail to accurately parse complex tables in multi-quarter financial disclosures.
Solution: Built a temporal RAG pipeline using markdown-based table parsing and metadata filtering to enable comparative financial analysis across fiscal periods.

📂 View Code

Foundational Research: High-Precision Instrumentation

PROJECT: CHARACTERIZATION & OPTIMIZATION OF TES DETECTOR SYSTEMS (U LETHBRIDGE)

Cryogenics Signal Processing NASA Testbed Interferometry

My foundational research at the University of Lethbridge focused on the far-infrared (far-IR) regime, a critical wavelength for interpreting cosmic origins. Working within the Astronomical Instrumentation Group (AIG), I specialized in the characterization of Transition-Edge-Sensor (TES) systems.

This research utilized a Double-Fourier-Interferometer (DFI) testbed, a collaborative development involving NASA Goddard and international partners. Modeled after the Herschel SPIRE mission architecture, the testbed employs a Michelson spatial interferometer coupled to a Mach-Zehnder-like Fourier Transform Spectrometer spectral interferometer. My work bridged the gap between cryogenic hardware and digital signal integrity.

DFI testbed @ U Lethbridge Instrumentation Lab

Key Engineering Contributions:

Signal Accuracy: Evaluated measurement precision by transitioning PID control logic from 32-bit integer to 32-bit float algorithms.
Noise Analysis: Derived and validated optimal noise parameters for off-axis pixels to ensure detector-array consistency.
System Calibration: Validated temperature calibration of feedback-controlled TES detectors across 1μA - 15μA current bias ranges.
Cross-talk Modeling: Analyzed thermal and electrical cross-talk experimental data to optimize multi-pixel isolation.
Efficacy Benchmarking: Compared rapid-scan vs. step-integrate approaches in FTS and DFI interferometry modes.