Projects

Investigating coverage‑guaranteed conformal prediction for cross‑agent uncertainty inheritance in dynamic DAG‑structured multi‑agent LLM systems - a gap unaddressed by existing frameworks (G‑Designer, DySCo, UProp, DebUnc). Identified and formalized theoretical flaws in an existing e‑process‑based topology design framework, including an independence violation and a "groupthink" failure mode from correlated agent errors. Developing calibrated confidence scoring between connected agents and a method to compose scores across a multi‑agent chain for a provable end‑to‑end accuracy guarantee.

Implemented a BERT‑style transformer text encoder from scratch in PyTorch (embeddings, multi‑head self‑attention, FFN, 12‑layer stack), validated to cosine similarity 1.0 against a reference bge‑base‑en‑v1.5 checkpoint, then fine‑tuned via contrastive learning (InfoNCE, hard + in‑batch negatives) on 82K mined query–passage pairs, improving Recall@1 0.39→0.65 and Recall@10 0.74→0.94. Architected a dual‑path multimodal retriever over a 638‑document PDF corpus, pairing the custom encoder with frozen CLIP for images, recovering ~1,200 figures via a CLIP‑based attachment gate at ~85% precision. Added cross‑encoder reranking (bge‑reranker‑base) for adaptive top‑k selection, grounding a Qwen‑VL backbone in retrieved text and figures to cut irrelevant context by ~30%, plus an observability layer logging per‑query retrieval paths, latency, and relevance scores.

Built a self‑attention Mixture‑of‑Experts recommender (MoSE), replacing LSTM experts with 4 parallel causal self‑attention experts under a shared MMoE gate for joint next‑item ranking and rating prediction on MovieLens. Achieved 0.282 Recall@10 / 0.163 NDCG@10 (~3× an LSTM‑expert baseline) while cutting rating RMSE from 1.177 to 1.097. Diagnosed emergent expert specialization via gate‑weight and attention analysis, consistent with MMoE routing dynamics reported in RecSys'19, and ran targeted ablations isolating an 11–12% Recall/NDCG gain from positional encoding.

Reviewed a paper on making reinforcement learning fair across multiple competing goals, framed as a game where one side pushes to improve the worst‑performing goal while the other tries to exploit any imbalance. Reimplemented both algorithms from scratch in Python, including a custom "dosa stall" simulation balancing revenue against customer rating. Found both algorithms broke down when goals were on very different scales, and fixed this by rescaling each goal to a common range before comparison, consistently improving fairness across all test scenarios.

Fine‑tuning an LLM for improved tool‑calling accuracy.

A cluster-based shard framework for machine unlearning in graph-based recommendation systems, using KMeans partitioning and adaptive aggregation. Cuts retraining cost by 85% while keeping accuracy within 1–2% of a full retrain - validated for user-, item-, and sample-wise unlearning. Published as "Machine Unlearning in Recommendation Systems," IEEE ICTEST 2025.

Implemented Graph Head Attention (GHA‑ViT) from scratch in PyTorch - replacing standard multi-head self-attention with Top-k sparse graph attention, as no public implementation existed - benchmarked against ViT on CIFAR-10. Reduced attention complexity from O(N²) to O(Nk), achieving 60.79% attention sparsity within 0.7% of baseline ViT accuracy (73.57% vs. 74.27%). Showed GHA-ViT's sparse attention consistently outperforming dense ViT under shuffle, noise, and masking perturbations. Designed two extensions - Multi-Scale k Graph Attention and spatial-bias attention - with the relative-position variant reaching 74.88% accuracy, beating both baselines.

Fine-tuned Google's gemma-3-1b-it (1B-param) for 6-class emotion classification across three low-resource, script-diverse languages - Santali (Ol Chiki), Kashmiri (Arabic script), and Manipuri (Meitei Mayek) - using QLoRA (4-bit NF4, rank=32, α=64) targeting attention and FFN projection layers, with no machine-translation bridge. Reformulated classification as instruction-tuned text generation and trained via SFT on 7,176 samples across a class-imbalanced 6-emotion label set. Ranked 7th in the competition with a Macro F1 of 0.51, under a single mandated 1B-parameter base model and a fixed evaluation protocol banning ensembling and external translation.

Built a ResNet-18 classifier trained entirely from scratch for 10-class flora/fauna image classification on 10K images, scoring 50% weighted F1. Designed a custom U-Net with a PixelShuffle-based 4× upsampler for joint low-light image denoising and super-resolution, achieving 37.7 dB best validation PSNR. Fine-tuned YOLOv8n for 6-class mosquito-species object detection, achieving 0.47 mAP@0.50 on validation.

Engineered a full-stack system with Vue.js 3 and a Flask REST API (15+ endpoints), implementing RBAC (Flask-Security) to separate librarian/member workflows and an approval-based issue-return system. Automated recurring operations with Celery + Redis task queues and Celery Beat - daily overdue reminders, monthly PDF/CSV report generation, and expired-reservation cleanup.

A multi-class classification model predicting recipe ratings, using Random Forest, SVM, and a Voting Classifier with heavy feature engineering on a Kaggle dataset. Cross-validated against F1-score and accuracy, landing a Top-50 leaderboard finish.