🤖 Ask My AI

Ask my AI about my research, projects, or how I think about problems.

✨ Featured Projects

Adaptive LLM Inference
A research implementation of adaptive inference for large language models, benchmarking early‑exit strategies on GPT‑2 Medium (SST‑2 sentiment classification) for accuracy–latency tradeoffs.
GPT‑2 Medium MLP Efficient LLM Inference
Adaptive inference comparison
Baseline vs. adaptive inference comparison
Confidence tradeoff
Confidence-based early exit tradeoff
MLP tradeoff
MLP early-exit accuracy vs. latency
Latency ↓ 74–81% vs Baseline GPT‑2
Accuracy Δ +13–14 pts vs Baseline GPT‑2
Speedup ~4–5× (0.021s → 0.004–0.005s)
  • Baseline GPT‑2 (no early‑exit, no fine‑tuning): 70.8% accuracy · 0.021s per sample.
  • GPT‑2 (confidence threshold early exit, no fine‑tuning): 49.1% accuracy · 0.00357s per sample.
  • GPT‑2 (confidence threshold early exit, fine‑tuning): 84.6% accuracy · 0.00403s per sample.
  • GPT‑2 (MLP‑based early exit, fine‑tuning): 85.1% accuracy · 0.00547s per sample.
Method Summary

Introduced confidence and MLP-based early-exit policies, evaluated across multiple checkpoints, and compared against full‑depth inference for latency/accuracy tradeoffs.

Impact Summary

Fine‑tuned early‑exit GPT‑2 boosts accuracy to 84.6–85.1% while cutting latency to 0.00403–0.00547s per sample (≈74–81% reduction, ~4–5× faster) compared to standard GPT‑2.

Shifting Stories — Shifting Returns
This study investigates how longitudinal changes in corporate narratives influence subsequent stock returns. Using Orbit quarterly reports and earnings call transcripts, it builds narrative features (sentiment, financial/business content, risk disclosure, subjectivity, temporal framing) and measures quarter‑over‑quarter shifts.
NLP Financial Text Quant Signals Backtests
Shifting returns summary slide
Project snapshot — narrative shifts & signal construction
MTS Method Mean Sharpe
MTS roll −8.37 −0.36
MTS w var −0.01 −0.16
MTS w freq −0.06 −0.29
MTS KL +0.03 0.46
MTS Cosine +0.12 0.54

Data Insights

  • Strongest narrative change (max |Δ|) carries the strongest predictive information.
  • Balanced directional accuracy ~54% Up and ~51% Down predictions.
  • Firms average ~15 feature changes per quarter; narratives are highly dynamic.
  • Higher future‑oriented firms saw ~30.4% returns vs ~7.1% for less forward‑looking firms.
  • Subjective (top quartile) ~50.8% returns vs Objective ~9.4%.

Result Highlights

  • Distance‑based measures (KL, cosine) delivered the strongest signals.
  • PCA blend of cosine + variance improved robustness and cumulative returns.
  • Factor regressions show positive alpha beyond market, value, and momentum.
Method Summary

Constructed Moving Target Scores (rolling, EWMA, median, variance, frequency, KL, cosine) to quantify quarter‑to‑quarter shifts. Distance‑based metrics produced the strongest signals, and a PCA composite of cosine + variance improved robustness.

Impact Summary

Signals based on KL divergence and cosine similarity delivered the most consistent performance. The best MTS variants reached Sharpe 0.54 with ~75% hit rates, and factor regressions showed positive alpha when narrative signals added information beyond traditional risk factors.

Fama‑French Factor
Replicated HML and SMB factor construction using CRSP/Compustat workflows, validating factor behavior and rolling alpha dynamics against published benchmarks.
HML SMB Factor Model Replication

Beyond replication, I analyzed zero‑dividend composition within value vs. growth and tracked its relationship to HML rolling alpha to understand when the value premium strengthens or weakens.

HML replication comparison
HML replication vs. benchmark
Rolling HML alpha
Rolling alpha dynamics (HML)
Zero-dividend market cap share within growth vs value and rolling alpha
Zero‑dividend share vs. HML rolling alpha
Method Summary

Constructed factor portfolios using standard HML/SMB definitions, validated time‑series behavior, and analyzed rolling alpha trends for value vs. growth regimes.

🎥 YouTube

DeepSeek’s mHC Explained: How It Improves LLMs

Stanford Alpaca: Revolutionizing AI

Understanding Self-Attention: The Core

✍️ Articles

Introduction to Risk Minimization

Introduction to Risk Minimization

 Polynomial Regression: Balancing Complexity and Generalization

Polynomial Regression: Balancing Complexity and Generalization

 What I found visualizing Covid-19 in Nepal?

What I found visualizing Covid-19 in Nepal?

🎓 College Projects

📦 LightResNet

A lightweight adaptation of the ResNet architecture from Microsoft Research designed for efficient image classification.

ResNet CNN PyTorch

🌍 Computing System Architecture

Implement single stage and five stage processor for RISC-V architecture.

Python RISC-V CPI

🎮 Machine Learning

Assessments on mathematical foundation behind machine learning. Used MATLAV for image analysis, regression and empirical risk.

MATLAB Image Analysis Risk Minimization