Building LLM-powered systems that retrieve, reason, and deliver grounded results.

Actively designing and scoping an AI voice agent for customer support that combines speech-to-text, LLM reasoning, RAG-based retrieval, and text-to-speech for natural, multi-turn phone interactions.

The MVP is being structured around FAQ resolution, order-status lookup, appointment scheduling, and human handoff for complex issues, with emphasis on latency, orchestration state, and reliability.

Current engineering focus includes targeting low-latency turn responses, managing state across multi-turn conversations, and designing interruption-safe flows for real-time voice interactions.

Built a large-scale customer complaint intelligence workflow using a financial complaints dataset with approximately 14M source records, framed as a multi-class NLP classification problem for complaint triaging.

Created a 500K sampled experimental setup and benchmarked TF-IDF + Logistic Regression, Linear SVM, and DistilBERT, achieving around 0.80 weighted F1 under practical constraints.

Current focus is on moving from notebook experimentation toward cleaner packaging, holdout-based evaluation, and more deployment-ready structuring for modular inference and future serving workflows.

Built an end-to-end multi-agent research system with 6 specialized agents (Planner, Search, Scraper, Retriever, Writer, Evaluator) to separate task planning, live source discovery, grounded synthesis, and output evaluation.

Designed a RAG pipeline using FAISS-based semantic retrieval over 5–20 chunks per query, processing 5–8 live external sources per run to improve grounding, traceability, and citation-backed generation.

Implemented an evaluation workflow across 50+ benchmark queries to assess relevance, faithfulness, and completeness, helping refine retrieval quality and improve the reliability of generated outputs.

Design tradeoff: modular multi-agent orchestration improved interpretability and evaluation control, while introducing higher latency than a simpler single-chain workflow.

Optimized the workflow to achieve approximately 15s end-to-end latency and structured the system for Docker-oriented execution and future CI integration.

Developed a document-grounded Q&A system over 10,000+ chunks using embedding-based retrieval with FAISS indexing.

Improved retrieval relevance from 65% to 82% through chunking refinement, embedding strategy improvements, and retrieval-quality iteration.

Reduced hallucination risk by strengthening retrieval grounding and designing the system with modular separation of retrieval, reasoning, and API-serving layers, alongside containerized deployment workflows.

Emphasized retrieval quality and modular backend design so document answers remained better grounded, easier to serve through APIs, and more reliable for downstream querying.

Key outcome: stronger retrieval quality and cleaner API separation made grounded document question answering more reliable.

Built an LLM-powered outreach automation system that extracts job descriptions, retrieves relevant project context, and generates personalized cold emails aligned to target roles.

Used ChromaDB + embeddings to improve retrieval precision for generation context, matching job requirements to relevant portfolio projects before response generation.

Designed the application with modular separation of ingestion, retrieval, and generation layers, and structured it for containerized execution with Docker.

Structured the system to support repeatable outreach generation with clearer separation between job parsing, project matching, and email generation, improving maintainability for future extensions.

Key outcome: modular retrieval-first design improved maintainability and made the workflow easier to extend with new targeting logic.

Built a credit risk prediction workflow to identify likely loan defaulters using structured financial data, with emphasis on preprocessing quality, feature selection, and interpretable risk modeling.

Cleaned messy data containing missing values, inconsistent entries, unrealistic records such as negative income values, and extreme outliers, then reduced 30–35 initial features to 20 meaningful variables through feature analysis, correlation review, and business relevance checks.

Applied SMOTE-Tomek to address class imbalance and compared results against a baseline modeling workflow, improving the robustness of default prediction on imbalanced financial data.

Improved model interpretability by removing weak predictors and surfacing less obvious variables that showed stronger association with default risk.