Sandbox for Building and Benchmarking Agentic AI Workflows

Added Jun 2025 3 design docs

Agentic AI systems are notoriously hard to evaluate: an agent that shines on one task fails quietly on another, and teams need somewhere to build, test, and compare agents rigorously before trusting them with real work. This project builds that environment. The intern creates an experimentation platform where agents built with CrewAI and Langchain, powered by OpenAI models, can be constructed, configured, and benchmarked side by side. The system supports defining test scenarios, running agents against them, and capturing results for comparison, turning subjective impressions of agent quality into measurable outcomes. Agent customization tools let users adjust prompts, tools, and collaboration patterns, while user profile management keeps each experimenter's configurations and history organized. A FastAPI backend orchestrates runs and a Streamlit frontend, all written in Python, makes experimentation interactive and repeatable. The project builds a skill the AI industry urgently needs, evaluation engineering: designing harnesses that reveal how agent systems actually behave. The intern finishes with both a working testing platform and demonstrated judgment about what makes agents reliable.

Related projects

Searchable Case Study Library for Banking Education

Case studies are one of the best ways to learn banking, since real events such as failed risk controls, successful product launches, and landmark regulatory actions teach lessons that textbooks flatten. The difficulty is discovery: relevant cases are scattered across sources, and a student rarely knows which one matches the topic at hand. The intern builds a streamlit application that makes a case study collection searchable by meaning rather than keyword. Case documents are embedded and stored in a vector database, and when a user asks about a topic such as credit risk in small business lending, a retrieval-augmented generation pipeline pulls the most relevant cases and generative AI produces a concise summary of each one, highlighting the key lessons and outcomes. Everything is written in python, and the interface lets users browse results, read summaries, and drill into the underlying case text. The project is an approachable introduction to RAG: the intern learns how embeddings capture topical similarity, how retrieval grounds generated summaries in real sources, and how to package the whole system as a clean educational tool for banking learners.

bankingresearchgenerative-aipython+4

View project

Medium

Batch Grading Analytics Pipeline for Student Assignments

A university processing tens of thousands of assignment submissions each term needs more than spreadsheets to compute grades, detect anomalies, and report cohort statistics. This project recreates that challenge as a distributed data workflow that an intern can build and run end to end. Submission events are simulated with Kafka so assignments arrive as a realistic stream throughout the exercise, and raw submission data is stored durably in HDFS. Scheduled PySpark jobs then process the accumulated submissions in batches, computing grading summaries, score distributions, completion rates, and per-course statistics. Python scripts manage the surrounding workflow, handling ingestion, triggering batch runs, and exporting results as structured data files and reports for instructors and administrators. There is no web or mobile front end; the deliverables are the pipeline itself and the reports it produces, which keeps attention on data correctness and throughput. The intern comes away understanding the streaming-to-batch pattern used across industry big-data platforms: durable ingestion, distributed computation, and reproducible reporting, exercised on a domain where the meaning of every metric is easy to reason about.

educationtechnologybig-datahdfs+3

View project

Medium

Retrieval-Backed Course Material Generator for Banking Education

Trainers and educators in financial services spend enormous effort producing course notes, quizzes, and explanations for every topic they teach, and the material dates quickly as products and regulations change. A generator that assembles fresh, accurate learning content from a maintained resource library changes the economics of banking education. The intern builds this generator in python with a streamlit interface. Educational resources, including articles, definitions, and worked examples, are embedded and stored in a vector database. When a user enters a banking topic or uploads their own learning materials, a retrieval-augmented generation pipeline gathers the most relevant source content and generative AI composes tailored outputs: explanatory lessons, topic summaries, and quiz questions with answers. Because generation is grounded through RAG in the stored resources, the produced material reflects the curated corpus rather than the model's unverified memory, and users can regenerate or refine content interactively. The project teaches content pipeline design, grounded generation for educational quality, and quiz construction logic, giving the intern a strong demonstration of AI applied to learning-content operations in a regulated domain.

bankingeducationgenerative-aipython+4

View project