I build and study systems that coordinate - whether that's autonomous agents sharing a task, or 238 countries trading with each other. My work sits at the intersection of agentic LLM systems and network science, and both come from the same instinct: that the structure of how things connect explains more than the properties of the things themselves.
I co-developed SW-KG, a multi-agent LLM coordination architecture that borrows small-world network topology to cut communication overhead between agents. My master's thesis modeled 35 years of UN Comtrade bilateral trade data as a geospatially embedded network and found that global trade is robust to random disruption but fragile to targeted removal of hub economies. Alongside the research, I've built production retrieval-augmented generation pipelines over enterprise data sources and designed database schemas and workflow architecture for a real-time tracking application.
I care about knowing what my results do and don't show - most of what I've learned came from the experiments that didn't work. Currently looking for applied scientist and AI engineer roles in generative AI, agentic systems, and network or supply chain intelligence.
Work Authorization: F-1 student with 12-month OPT + 24-month STEM OPT eligibility (up to 3 years of work authorization, no employer sponsorship required during that period).
91.8% token reduction in multi-agent LLM coordination (SW-KG)
35 years of global trade modeled as an evolving network
Published researcher · peer-reviewed journal + preprint
Co-developed a multi-agent LLM coordination architecture that arranges agents in a
Watts–Strogatz small-world graph (K=4, p=0.3) rather than a fully connected mesh,
layered with a token economy and a hub-promotion mechanism. Measured 91.8% token
reduction and 12.19× latency speedup at 10×100 agent scale, validated through
controlled ablation studies isolating each component. Published as a preprint.
PythonLangChainNetworkXMulti-Agent SystemsAgentic AI
Modeled 35 years of UN Comtrade data (1988–2022) as small-world supply chain
networks (238 countries, 16,000+ links). Confirmed 7 of 9 hypotheses (2 partial),
including ~80% fragmentation under targeted hub removal and Asia-Pacific regional
clustering with C = 0.91.
Data-driven analysis of Linux CPU scheduling (CFS) by collecting real process
metrics with top, ps, and pidstat on Ubuntu. Compared real scheduler behavior
and CPU utilization patterns against textbook algorithms using Python-based
analytics and visualizations.
University Database Design (Relational Data Modeling)
Designed and implemented a fully normalized relational database for a
university/company scenario, modeling entities such as students/employees,
departments, courses/projects, and enrollments/works_on. Built ER diagrams,
enforced integrity constraints, and wrote complex SQL queries, views, and
stored procedures for realistic reporting and analytics use cases.
Applied K-Means and DBSCAN clustering with SARIMA time-series forecasting to
28,000+ global seismic records (1960–2023). Identified five clusters aligning
with major tectonic boundaries; magnitude forecasting reached 5.2% MAPE on a
held-out test period. Published in a peer-reviewed journal.
Multi-agent LLM systems spend most of their tokens talking to each other. SW-KG addresses this structurally rather than through prompt engineering — by changing the topology agents communicate over. Agents are arranged in a Watts–Strogatz small-world graph (K=4, p=0.3), which preserves short average path length between any two nodes while using far fewer edges than a complete graph. A token economy meters communication cost, and a hub-promotion mechanism elevates frequently-routed nodes.
Key Results:
91.8% reduction in token usage against baseline at 10×100 agent scale
12.19× latency speedup at the same scale
Each architectural component isolated through controlled ablation studies
Limitations & Current Work:
Hub promotion did not carry its weight on token metrics in isolation — it improved routing behavior but cost more than it saved in tested configurations
The theoretical complexity bound in the preprint is stronger than the empirical results at tested scales support
Current traversal does not fully exploit the shortcut edges the topology provides
A v2 is in progress exploring a hybrid design pairing HNSW-based retrieval with SW-KG's coordination layer
Small-World Spatial Network Analysis of Global Supply Chains Using International Trade Data (1988–2022)
Modeled global supply chains as geospatially embedded small-world networks, integrating network science and spatial analysis across 238 reporting economies and 16,000+ trade links. Key findings: pronounced small-world structure, strong regional clustering, and asymmetric resilience under targeted disruptions. Manuscript and code repository will be linked after submission.
Key Insights:
Global supply chains show strong small-world properties (σ > 1.2 across all intervals) with high clustering and short path lengths
Networks are resilient to random failures but vulnerable to targeted hub removal — ~15–20% hub removal fragments ~80% of the network
Regional clusters (EU, East Asia, North America) have high internal clustering, with Asia-Pacific reaching C = 0.91
Geography and clustering are strongly correlated (r = 0.72)
Cross-regional corridors are critical for efficiency and global connectivity
Earthquake Pattern Analysis Using Clustering, Forecasting & Machine Learning: A Global Study (1960–2023)
Published · Journal of Research in Environmental and Earth Sciences (QUEST Journals), Vol. 11 Issue 6 | Read Paper
A global study of seismic activity from 1960 to 2023, combining unsupervised clustering, time-series forecasting, and supervised risk classification over 28,000+ earthquake records from the USGS National Earthquake Information Center catalog.
Results:
K-Means (K=5, selected via elbow method) segmented events into five clusters aligning with major tectonic boundaries — the Pacific Ring of Fire, Himalayan Belt, Mid-Atlantic Ridge, and San Andreas Fault system
DBSCAN (eps=0.5, min_samples=10) isolated dense seismic zones and separated outlier events from clustered activity
SARIMA magnitude forecasting achieved 5.2% MAPE on a 2016–2023 holdout, with residual analysis showing minimal bias
Random Forest risk classification reached 72.2% accuracy (74.1% precision, 70.8% recall); a Decision Tree baseline reached 67.5%
A Note on Data Provenance:
The location-probability ranking surfaced a limitation worth documenting: several of the highest-probability sites correspond to former nuclear test locations — the Nevada Test Site and Semipalatinsk in Kazakhstan — rather than to natural seismicity. The source catalog does not distinguish anthropogenic events from tectonic ones, and no filter was applied. A revised version addressing this is planned.
Interactive Visualization
Earthquake Data Heatmap: interactive visualization of global earthquake density and clustering, used to identify seismic hotspots across the 1960–2023 catalog.
Work Experience
AI Engineer Intern
Gogentic AITexas, USA (Remote)June 2025 – August 2025
Built retrieval-augmented generation (RAG) pipelines in Python with FAISS vector indexing over NeuroVault neuroimaging datasets, enabling semantic search across unstructured scientific data
Unified three enterprise data sources (Oracle, SQL Server, PostgreSQL) into a single retrieval layer and automated LLM-driven model workflows for downstream analysis
Developed AI meeting-summarization tooling with real-time voice-to-text transcription and privacy-aware analytics
Web Application Architect
Studium SpanMadhya Pradesh, IndiaJul 2023 – Jul 2024
Designed the database schema and workflow architecture for a real-time construction progress tracking application, and led the team that built it
Designed and developed SQL-based reports and interactive dashboards using Power BI and Tableau, enhancing operational efficiency by 20%
Automated data scrubbing and validation processes, improving system accuracy by 15%
Collaborated with cross-functional teams to troubleshoot technical problems and ensure seamless workflows
Web Application Architect Intern
Studium SpanMadhya Pradesh, IndiaJan 2023 – Jun 2023
Implemented data integration solutions and built predictive analytics tools using Python and SQL
Utilized Power BI to visualize complex data, driving actionable insights for business decisions
Authored detailed process documentation to align with data governance standards
Education
Master of Science, Data Science
University of Massachusetts, Dartmouth2024 – 2026 · Thesis defended May 2026 · Degree conferring Aug 2026
Relevant Coursework:
High-Performance Scientific ComputingAdvanced Mathematical StatisticsSmall World NetworksAdvanced Data MiningSoftware Testing and AutomationBusiness Intelligence and Data MiningDatabase DesignOperating Systems
Bachelor of Technology, Computer Science Engineering
Cloud ComputingDatabase Management SystemsData Mining and WarehousingMachine LearningComputer NetworksAnalysis Design of AlgorithmObject Oriented ProgrammingInternet of Things
