Selected work
Applied work spanning RAG, LLM agents, graph retrieval, forecasting, and NLP — drawn from research, industry roles and public projects. Currently building in public: a ReAct agent from scratch, a LangChain rebuild and a GraphRAG project (github.com/sadrjahani).
Local PDF RAG System
A fully local question-answering pipeline over PDF books — PyMuPDF, FAISS and an Ollama model — with source citations and a 68-question retrieval evaluation.
View projectReAct Agent From Scratch
A ReAct agent in plain Python, without agent frameworks. Switches between Anthropic, OpenAI and local Ollama models, runs a Thought → Action → Observation loop and handles malformed output and tool errors. In progress.
View projectNeo4j Semantic Search for LLM Apps
Consent-based data collection on GCP feeding a Neo4j graph that powers semantic search for LLM applications, built at Goodfolio.
View projectSiamese BERT Sentence Scorer
An autoencoder and Siamese BERT model scoring semantic textual similarity between augmented sentence pairs, benchmarked against SBERT and Augmented SBERT.
View projectDynamic Pricing Optimizer
A pricing system overhaul that improved efficiency by 30% while cutting execution time, built freelance.
View projectTraffic Flow Forecasting
Predictive models comparing LSTM, XGBoost, ARIMA, and Prophet for vehicular traffic-flow prediction.
View projectFraud & Sentiment Detection
NLP-driven fraud detection and sentiment analysis models delivered in both English and Persian.
View projectE-Commerce Pipeline Automation
Automated data pipelines and marketing-data optimization for an e-commerce platform's decision-making.
View projectMarket Intelligence Web Scraping
Automated collection of competitor and market data to feed pricing and demand models, built freelance.
View projectDecision-Making Dashboard
A management dashboard consolidating network and infrastructure health for faster operational decisions, built at ASIGI.
View project