TechnicalPortfolio

Architected a stateful system using LangGraph and SQLite, ensuring 100% context retention across multi-turn sessions.

Optimized inference by integrating Groq (LLaMA 3.1) to achieve 560 token/s, enabling near-instant fact-based decision support.

Engineered an autonomous Tool-Execution layer with live web-search for real-time fact validation.

Architected a Dual-Path Hybrid Retrieval engine using EnsembleRetriever, fusing Dense Vector embeddings for semantic context with Sparse BM25 matching for lexical precision.

Optimized retrieval by implementing a custom weighting ratio, ensuring the system captures both high-level meaning and exact-match terminology.

Developed a Multi-Modal Ingestion Pipeline to generate text-based summaries of visual PDF elements (tables/charts), ensuring 100% data discoverability for non-textual content.

Developed a custom Convolutional Vision Transformer and Deformable DenseNet specifically for remote sensing data.

Achieved a 37% reduction in classification error by optimizing the model for satellite images.

Published in IEEE, proving the model's viability for large-scale environmental monitoring.

Architecting a Dockerized Microservices environment for deployment on AWS, ensuring high availability and horizontal scaling.

Implementing Celery and Redis to manage heavy document ingestion tasks asynchronously, keeping the main application responsive.

Integrating Supabase (PostgreSQL) with pgvector for high-speed hybrid retrieval, and Clerk for robust user authentication.

Proposed a Backpropagation-free training framework utilizing closed-form, layer-wise optimization.

Implemented a local learning approach that bypasses global gradient updates, reducing the overall computational footprint.

Achieved up to 76% lower error rates and superior convergence speed compared to standard Adam/AdamW optimizers on benchmark datasets.

Designed a Geospatial Graph Topology representing census tracts as nodes and spatial adjacency (contiguity) as edges, enabling the model to explicitly learn from regional "spillover" effects.

Built a Graph Convolutional Neural Network (GCN) to fuse regional socioeconomic variables with the underlying graph structure, capturing latent spatial dependencies.

Achieved a 10% improvement in predictive accuracy over non-graph baselines (e.g., SVR) by leveraging the Spatial Adjacency Matrix to inform predictions.

Built a Reflexion-based Agent using LangGraph that enables the model to critique, self-evaluate, and refine its own output iteratively.

Integrated TavilySearch API for autonomous web retrieval, enabling the agent to cross-reference facts before finalizing content.

Developed a multi-step reasoning chain that increased content accuracy for complex, fact-heavy prompts.