SYSTEM DEPLOYMENTS

DATA PIPELINES_

NoteSummarizer

#01

Built an end-to-end AI agent that transforms messy construction site notes into structured weekly briefs using a two-pass LLM pipeline (extraction → synthesis) with Gemini 2.5 Flash and automatic Claude failover — proven against a real API outage. Features multi-project segregation, uncertainty flagging, sensitive-content guardrails, and a Streamlit chat-style UI with file attachment and drafted-email delivery. Supports text, PDF, and combined inputs. Found and fixed three real prompt-engineering bugs through adversarial stress testing, each diagnosed to root cause.

STACK CONFIGURATION:

Anthropic Claude Sonnet · Streamlit · pdfplumber · conda · GitHub

HireLens AI

#02

HireLens AI is a full-stack AI hiring platform where recruiters post jobs and candidates apply via a shareable link — their resume is automatically screened by a Gemini/Qwen/DeepSeek ensemble, and those who pass complete a hands-free avatar interview driven by real-time silence detection. Every answer is recorded and stored; recruiters later evaluate transcripts against a weighted rubric and cross-check resume claims against what candidates actually said. Key decisions: a 3-model ensemble after observing ±29% single-model scoring variance; silence-based turn-detection after live Whisper transcription proved too slow on CPU. Built with Python, Streamlit, WebRTC, PyAV, FFmpeg, Whisper, and SQLite.

STACK CONFIGURATION:

Streamlit · Gemini · Qwen · DeepSeek · OpenRouter · Whisper · gTTS · streamlit-webrtc · PyAV · FFmpeg · SQLite · OpenCV

Credit Risk Prediction Model

#03

Developed an end-to-end credit risk classification system on 150,000 borrower records using XGBoost and Random Forest ensemble models. Applied median imputation, outlier capping, and engineered the TotalTimesLate feature by aggregating multi-horizon delinquency signals — achieving an AUC-ROC of 0.8692. Demonstrated production-ready ML pipeline design with direct applicability to fintech credit underwriting and risk scoring workflows.

STACK CONFIGURATION:

ML · Python · XGBoost · Scikit-learn

FinRAG-AI: Financial Document Chatbot

#04

Architected a production-grade RAG pipeline for 10-K filings and financial reports, combining a local FAISS vector store with Google Gemini 2.5 Flash for hybrid retrieval-generation. Implemented RecursiveCharacterTextSplitter chunking, HuggingFace all-MiniLM-L6-v2 embeddings, and a LangChain LCEL inference chain — delivering contextual financial Q&A via a Streamlit interface with sub-second retrieval on multi-page regulatory documents.

STACK CONFIGURATION:

RAG · LangChain · FAISS · Gemini 2.5

Stock Price Forecasting with LSTM

#05

Designed a PyTorch LSTM network for stock price forecasting using a 60-day sliding window on AAPL historical data sourced via yfinance. Executed hyperparameter tuning — scaling hidden layers to 128 units across 30 epochs — reducing RMSE from $13.12 to $11.57 on a strict 80/20 chronological train-test split. Demonstrated rigorous time-series methodology with leakage prevention and directional accuracy benchmarking.

STACK CONFIGURATION:

Deep Learning · PyTorch · Time Series

FinSense: Financial NLP Sentiment Engine

#06

Built a high-performance NLP pipeline using FinBERT (finance-domain pre-trained transformer) and Meta FAISS for semantic financial news retrieval. Encoded the Financial PhraseBank into 768-dimension vectors with L2-normalized cosine similarity search, implemented mini-batch embedding for CPU-stable inference, and delivered sub-1ms query response — enabling context-aware retrieval across semantically related financial concepts without keyword dependency.

STACK CONFIGURATION:

NLP · FinBERT · FAISS · Transformers

Bank Marketing Campaign Analysis

#07

Analyzed 45,211 customer records from UCI's Bank Marketing dataset to diagnose low term-deposit conversion rates and optimize outbound campaign strategy. Segmented customers by demographics and contact behavior using Python and SQL KPI logic, uncovering that prior-success customers converted at 65% versus an 11.7% baseline — delivering actionable targeting recommendations and a Power BI dashboard to support data-driven campaign execution.

STACK CONFIGURATION:

Analytics · Python · SQL · Power BI