( Archive )

Projects.

Research systems, applied ML builds, and hackathon prototypes, most with code on GitHub.

Projects
13
Code
GitHub ↗

( 01 — Everything )

01 LLM Evaluation 2026
16K context sweeps · 5 adversarial families

Long-Context Adversarial Eval Harness

  • Built a uv-based evaluation harness for source-grounded long-context fact QA, shaped after RULER, BABILong, and LongBench: configurable 2K/8K/16K length buckets and early/middle/late evidence positions.
  • Hand-verified 60-case synthetic dataset across five adversarial families (single fact, multi-key distractor, multi-value, multi-hop, misleading fact pair), with an audit log checking every answer against an exact source span.
  • Forced strict JSON with cited evidence spans so hallucinations are auditable; reports accuracy, schema validity, evidence-span presence, and calibration (ECE with reliability diagrams) per family and bucket.
  • Ran GPU inference on Modal with vLLM and added a regression command that fails CI when a previously passing case breaks; a 0.5B Qwen2.5 model scored 0.167 accuracy at up to 16K, exposing sharp long-context degradation.
vLLM · Modal · uv · pytest Code
02 Civic ML 2026
A* road-aware patrol planning

ParkWatch

  • Built an AI-powered parking-enforcement intelligence dashboard for Bengaluru traffic teams with GraphSAGE hotspot forecasting and Mappls road-aware A* patrol sequencing.
  • Added traffic-delay exposure estimates, analyst explanations, scenario analysis, and exportable action plans through a Next.js + FastAPI stack.
Next.js · FastAPI · GraphSAGE · Mappls Code
03 Research Systems 2026
100K+ arXiv papers indexed

Research Corpus Agent

  • Designed a multi-agent RAG system for scientific literature analysis over 100K+ arXiv papers using planner, retriever, analyst, and critic agents.
  • Constructed a hybrid retrieval stack with BGE, BM25, ColBERT, reciprocal-rank fusion, cross-encoder reranking, and multimodal figure/table understanding.
BGE · BM25 · ColBERT · RRF Code
04 Signal ML 2026
0.971 avg. F1, 5-fold CV

GNSS Anti-Spoofing Detection

  • Designed an early-fusion Transformer for GNSS spoofing detection over 112-dimensional multi-channel feature sequences.
  • Built physics-informed feature engineering pipelines and achieved a 0.971 average F1-score across 5-fold cross-validation.
Transformers · Feature engineering
05 Tabular ML 2026
0.84 ROC-AUC with XGBoost

Credit Card Default Prediction

  • Developed an end-to-end credit-risk prediction pipeline on 30K customer records with SMOTE-based imbalance handling and model comparison across Logistic Regression, Random Forest, LightGBM, and XGBoost.
  • Performed Bayesian hyperparameter optimization with Optuna and SHAP-based explainability, reaching 0.84 ROC-AUC with XGBoost.
XGBoost · LightGBM · Optuna · SHAP Code
06 Computer Vision 2025
+10pt mAP@0.5 (35% → 45%)

GATE-IR

  • Formulated a gated thermal perception pipeline that routes infrared images through weather-specific enhancement modules before YOLO-based detection.
  • Improved thermal detection performance from 35% to 45% mAP@0.5 while preserving near-YOLO latency.
Thermal imaging · YOLO · Gating Code
07 Diffusion LLMs 2025
78% GSM8K, up from ~60%

ASMS: Adaptive Semantic-Momentum Sampling

  • Proposed a diffusion LLM decoding framework combining confidence momentum, semantic hysteresis, and entropy-aware control to mitigate token oscillation and hallucination lock-in.
  • Improved GSM8K reasoning accuracy from roughly 60% to 78% over confidence-based remasking baselines.
dLLMs · Remasking · Decoding Code
08 Efficient Inference 2025
1.7× decoding speedup, 97% QA retained

FOCUS: Sparse Token Selection

  • Trained a lightweight neural token-pruning model on SQuAD attention distributions to identify high-utility context tokens before inference.
  • Pruned candidate context tokens by over 60% and achieved up to 1.7x decoding speedup while retaining over 97% of dense-attention QA performance.
Token pruning · Selective attention
09 Generative Tools 2025

AutoGenx

  • Built a multimodal website-generation platform using Flask and Gemini APIs to transform text prompts, sketches, mockups, and Figma wireframes into deployable web apps.
  • Implemented code generation, packaging, live preview, and GitHub Pages deployment workflows.
Flask · Gemini Code
10 Health Tech 2025

Elyx Health Dashboard

  • Built a full-stack health intelligence platform with Flask, JavaScript, Tailwind CSS, and Chart.js for biomarkers, wearables, diagnostics, and patient-engagement views.
  • Integrated GPT-based retrieval and question answering over structured and unstructured healthcare data.
Flask · Tailwind · Chart.js Code
11 Social Prediction 2025

ViralStake

  • Developed a prediction platform using FastAPI, MongoDB, Streamlit, and Twitter APIs, enabling users to stake virtual assets on tweet-virality forecasts.
  • Integrated Gemini-powered tweet analysis to generate virality scores and explainable engagement insights.
FastAPI · MongoDB · Streamlit Code
12 Robust Vision 2025
96% accuracy, CLIP ensemble

Adversarially Attacked Image Classification

  • Designed a Continuous Attribute Learning architecture using ResNet50 and multi-task learning to study interpretability, overfitting, and failure modes.
  • Built a CLIP ensemble with prompt ensembling and test-time augmentation, achieving 96% accuracy and outperforming a single-model zero-shot baseline by 6%.
ResNet50 · CLIP · TTA Code
13 Multimodal Generation 2024

Multimodal Story Generation

  • Fine-tuned OPT-1.3B using LoRA/PEFT on fantasy storytelling data and integrated it with Stable Diffusion 1.5 for a multimodal text-image generation pipeline.
  • Implemented coherence scoring, dynamic summarization, entity tracking, and adaptive parameter control for long-horizon narrative consistency.
OPT-1.3B · LoRA · SD 1.5 Code