Los Angeles, California Est. 2026
The Daily Report

Personal Portfolio

Maitreyi Pareek

The Personal Record of a Student

July 24, 2026 · Vol. I · Selected Works · Notes
Selected Works

The Projects Desk

A field report on things I've built — from agentic AI systems to the file system underneath them.

Reflective Hybrid RAG for Recommendations
A

Reflective Hybrid RAG for Recommendations

Agentic AI Project

How can a recommendation system separate what a user has already done from what they actually want next — and use that gap to personalize better?

Improved top-1 recommendation accuracy 3× (12.5% → 37.5% HR@1) with a reflective hybrid RAG system separating raw user history from synthesized preference memory. Lifted personalization quality across ranking metrics (+25% HR@10) by combining dense embeddings and sparse SPLADE retrieval with LLM-generated user personas.

RAG LLM Agents SPLADE Recommender Systems
Problem, Limitations & Findings

The Problem

How can a recommendation system separate what a user has already done from what they actually want next — and use that gap to personalize better?

Limitations

Relies on LLM-generated personas, which can drift or hallucinate preferences the user never actually expressed. Evaluated on offline benchmarks, not live user traffic.

Findings

Separating raw history from synthesized preference memory tripled top-1 accuracy (12.5% → 37.5% HR@1) and lifted HR@10 by 25%; combining dense embeddings with sparse SPLADE retrieval outperformed either approach alone.

GitHub ↗ Paper ↗
Adaptive Subset Selection for Federated Learning
B

Adaptive Subset Selection for Federated Learning

FedAvg, FedProx

How can we create a subset of the bigger data that mimics the actual data of the server, and how can we modify the model per client when each device works differently?

Cut federated training cost 70% per client via adaptive coreset selection while preserving accuracy under FedAvg on non-IID CIFAR-100 partitions. Built a reproducible federated evaluation framework across 20 training rounds using ResNet-18 with client-side subset selection under realistic non-IID constraints.

Federated Learning ResNet-18 CIFAR-100 Adaptive Coreset Selection
Problem, Limitations & Findings

The Problem

How can we create a subset of the bigger data that mimics the actual data of the server, and how can we modify the model per client when each device works differently?

Limitations

Tested on CIFAR-100 under simulated non-IID partitions, not real edge devices; coreset selection adds client-side compute overhead that could matter on low-power hardware.

Findings

Adaptive coreset selection cut per-client training cost by 70% across 20 federated rounds while preserving accuracy under FedAvg, even under realistic non-IID data splits.

GitHub ↗ Paper ↗
Virtual File System
C

Virtual File System

Systems Programming

What does an operating system actually have to track just to open, read, and write a single file?

Engineered a virtual file system in Java with an I/O subsystem, open file table, bitmap-based block allocation, and file descriptors, enabling directory management, metadata tracking, and efficient block-level file access.

Java Operating Systems File Systems
Problem, Limitations & Findings

The Problem

What does an operating system actually have to track just to open, read, and write a single file?

Limitations

Built as a simulation on top of a virtual disk image, not integrated with a real OS kernel or tested under concurrent multi-process access.

Findings

A working I/O subsystem, open file table, bitmap-based block allocator, and file descriptor layer together were enough to support directory management and efficient block-level access.

GitHub ↗
D
D

Search Engine, Information Retrieval System

Information Retrieval

How do you rank the right answer out of 37,000 documents instead of just the one with the most matching words?

Built a Python search engine over 37,000 documents using a custom inverted index and TF-IDF ranking, improving retrieval with stemming, term weighting, duplicate detection, and multi-term query support.

Python TF-IDF Inverted Index IR
Problem, Limitations & Findings

The Problem

How do you rank the right answer out of 37,000 documents instead of just the one with the most matching words?

Limitations

TF-IDF ranking doesn't capture semantic meaning the way modern embedding-based retrieval does; evaluated on a fixed document set rather than continuously updated data.

Findings

A custom inverted index with stemming, term weighting, and duplicate detection made multi-term queries fast and relevant across the full 37,000-document corpus.

GitHub ↗
E
E

LLM Confidence Calibration

LLM Evaluation

Can you trust an LLM's confidence score, or is it just as confidently wrong as it is right?

Evaluated Llama 3.3-70B's self-reported confidence across five benchmarks (including GSM8K and MMLU), finding that 96–97% of incorrect answers were still scored with 80%+ confidence — surfacing a sharp overconfidence gap relevant to safe LLM deployment.

LLM Evaluation Llama 3.3-70B GSM8K MMLU
Problem, Limitations & Findings

The Problem

Can you trust an LLM's confidence score, or is it just as confidently wrong as it is right?

Limitations

Findings are specific to Llama 3.3-70B and the five benchmarks tested; confidence was self-reported by the model rather than derived from internal logits or probabilities.

Findings

96–97% of incorrect answers were still reported with 80%+ confidence, showing a sharp overconfidence gap that matters for any system that trusts an LLM's self-assessment.

GitHub ↗
F
F

Study Assistant Agent

AI Agents

Can an AI agent actually help you study, or does it just answer questions without reinforcing what you're supposed to learn?

Built a command-line AI study assistant using LangChain and LangGraph with tools for quiz generation, persistent note storage, and secure self-evaluation, powered by Groq-hosted Llama 3.3.

LangChain LangGraph Groq Llama 3.3
Problem, Limitations & Findings

The Problem

Can an AI agent actually help you study, or does it just answer questions without reinforcing what you're supposed to learn?

Limitations

Runs as a command-line tool rather than a full app; quiz and evaluation quality depend on the underlying model's output rather than a fixed curriculum.

Findings

Combining LangChain and LangGraph tools for quiz generation, note storage, and self-evaluation produced a working end-to-end study loop powered by Groq-hosted Llama 3.3.

GitHub ↗