VK
Vishwas Kothari
CS Graduate Student & AI Researcher

Vishwas Kothari

I build things that are meant to be understood, not just by machines, but by the people who depend on them.

My background spans two worlds: applied research at ISRO's Space Applications Centre, where I built optimization models to extract meaningful signal from satellite sensor data, and rigorous graduate-level ML at the University of Colorado Boulder, where the focus is less on what models do and more on why they work.

Along the way I've published in Springer, earned an AWS certification, peer-reviewed for Elsevier, and built end-to-end ML systems that prioritize explainability and accountability because a model no one can audit is a model no one should trust.

Outside work: cricket, badminton, and a reading habit that leans toward technology, business, and the occasional uncomfortable idea.

Explainable AI

Building models that earn trust through transparency: SHAP, LIME, DiCE, and fairness auditing baked into every pipeline.

Data Engineering

End-to-end data pipelines at scale, from satellite sensor data to financial documents, with validation at every step.

Machine Learning

PyTorch, TensorFlow, XGBoost, and vision-language models applied to problems where the stakes are high enough that getting it wrong actually matters.

Research Engineering

Turning research ideas into reproducible experiments, clear reports, and usable demos with Python, AWS fundamentals, Git, Docker, Streamlit, and Gradio.

Jan - Jul 2025

Research Intern

Space Applications Centre — ISRO · Ahmedabad, India

  • Built end-to-end Python pipelines (SciPy, NumPy, xarray) for satellite data processing & NetCDF merging
  • Developed Levenberg-Marquardt optimization for bio-optical parameter retrieval (R²=0.987)
  • Created cross-sensor validation framework with 6 plot types & 4 statistical metrics
  • Validated chlorophyll-a retrieval (log-space R²=0.879) and authored 60+ page technical report
2025 - 2027

Professional Master's in CS

University of Colorado Boulder · GPA 3.5/4.0

Graduate-level ML with focus on model interpretability and real-world deployment.

2021 - 2025

B.Tech in Information Technology

Vellore Institute of Technology · GPA 8.49/10.0

Foundation in computer science, algorithms, and software engineering.

FinAgent-Eval

Qwen2.5-VLPix2StructDonutLayoutLMv3StreamlitSEC Filings

Reproducible evaluation harness for visual question answering over real SEC filing screenshots. Hand-curated 397 questions from 10 US public companies across extractive, layout, numerical reasoning, and chart interpretation tasks, then compared 5 models: Pix2Struct, Donut, LayoutLMv3, OCR+RoBERTa, and Qwen2.5-VL-7B on a free Colab T4. Qwen delivered a 4.3x ANLS lift, but numerical reasoning remained unsolved with 0.1138 ANLS and 0% paraphrase consistency.

View on GitHub

Financial Document VQA

PyTorchLayoutLMv3DonutPix2StructLoRAGradio

Benchmarked 4 vision-language models on SEC 10-K filings, revealing 63-83% performance drop vs generic DocVQA. Domain-specific LoRA fine-tuning on 307 examples improved ANLS by 22%, while generic fine-tuning caused 39% negative transfer.

View on GitHub

XAI Credit Lens

PythonSHAPLIMEDiCEXGBoostOptunaStreamlit

End-to-end explainable credit risk framework: 0.782 AUC on 30K records with automated fairness auditing and zero violations across four protected attributes. Regulatory compliance mapping for ECOA, EU AI Act, and Fed SR 11-7.

View on GitHub

Implicit Bias of Adam vs. SGD

PyTorchCVXPYNumPyMatplotlibMargin TheoryOOD

Investigated why different optimizers find fundamentally different solutions in overparameterized models. Connected recent theory showing GD converges to the ℓ₂ max-margin solution while Adam converges to ℓ∞, then verified this empirically across 5 seeds with >0.98 cosine similarity to exact convex-programming solutions. Demonstrated practical consequences on an MNIST spurious-correlation benchmark: SGD suffered a 30% OOD accuracy drop and 3× higher shortcut reliance, while Adam reduced the drop to 13%.

View on GitHub

XAI-Driven Collision Avoidance

PythonSHAPLIMERandom ForestExplainable AI

Interpretable collision detection system for V2V/V2I communication in the Internet of Vehicles. Achieved 99.9% AUC on 206K instances with transparent feature attribution via SHAP and LIME for safety-critical AI decision-making.

Manuscript in Preparation

AWS Certified Cloud Practitioner

Score: 970/1000 (2024), deep fluency in AWS core services and cloud architecture.

Springer Publication

Kothari, V., Krishwanth, B. (2025). Empowering Survivors: Ethical AI for Countering Violence Against Women.

DOI: 10.1007/978-981-96-6046-9_26

Peer Reviewer

Public Health and Social Sciences and Humanities Open (Elsevier Journals, 2026-Present)

Leadership

Senior Core Committee: E-Cell, VIT-Stellar, VIT-Spartans, Anokha NGO, Youth Red Cross (2023-2024)

PythonSQLJavaC/C++JavaScript PyTorchTensorFlowXGBoostLightGBMscikit-learn SHAPLIMEDiCEOptunaLLM Workflows NumPyPandasSciPyxarrayNetCDFOpenCV AWSDockerGitLinuxStreamlitGradioMySQLMongoDBPower BI

Let's build something
meaningful

vishwasvkothari@gmail.com

(720) 388-4856