I build machine learning systems that can explain themselves — and the interfaces that make those explanations useful to someone who isn't an ML engineer.

I'm a Computer Science undergrad at VIT Chennai (AI & ML specialisation, class of 2027). I spent a summer at the Indira Gandhi Centre for Atomic Research replacing a manual industrial weld-inspection workflow with an end-to-end YOLOv8 pipeline, and I've presented a paper on automated AWS cost and security cleanup at ICANDIT 2026.

The through-line across everything below is the same, whether it's a research notebook or a game engine: a system nobody can inspect is a system nobody should trust. So my projects ship with the model and the dashboard, the metric and the interface, the paper and the CI pipeline.

Most ATS tools hand you a score. This one hands you the reasoning — Shapley values, LIME, counterfactuals and attention heatmaps that name the exact skills that earned the score, and price out what adding the missing ones would be worth. Built on a 22-skill / 346-alias ontology and SBERT embeddings over 1M+ job postings and 2,484 real resumes, with the Shapley and LIME implementations written from first principles rather than imported. The 6 GB of model artifacts stream from memory-mapped arrays, which is what makes it deployable at all.

Deployed live and usable in a browser.

FastAPI Streamlit sentence-transformers scikit-learn Plotly Docker Kubernetes Nginx

Live demo  ·  Datasets  ·  Models

An adversarial robustness study of post-hoc explainability: can an attacker leave a security model's prediction untouched while corrupting the explanation the analyst reads? Across 235,795 phishing URLs, an intrusion-detection set and a fraud corpus, four attacks against SHAP, LIME and Integrated Gradients say yes. Four defenses follow — the hybrid one cuts mean explanation drift by 91.1% while holding classification performance. A few-shot analysis shows the problem gets 40–100% worse in the low-data regime, which is exactly where security teams operate.

PyTorch SHAP LIME Captum XGBoost scikit-learn pytest

Symptom-to-diagnosis prediction across 41 conditions and 133 symptoms, wrapped in explainability at every layer: calibrated confidence, a ranked differential, a FAISS-retrieved evidence paragraph, and an optional LLM explanation in plain English. Seven classifiers benchmarked under 5-fold CV. If FAISS or a GPU isn't available it degrades gracefully to TF-IDF retrieval, so it runs on any laptop — and the LLM layer is optional by design, never load-bearing.

The front end is a dark-themed responsive Flask UI with live symptom search, chip selection and an animated confidence bar. No build step, no bundler.

Flask scikit-learn FAISS sentence-transformers Jinja2 Vanilla JS pytest GitHub Actions

Cloud accounts accumulate ghosts: forgotten EC2 instances, orphaned EBS snapshots, stale IAM users, security groups nobody remembers opening. This scans seven AWS resource domains, pulls real CloudWatch utilisation, and runs a Random Forest over engineered idle-signal features to recommend keep / review / delete with a confidence score — falling back to a rule engine when no model is present.

Deletion is treated as the dangerous operation it is: dry-run by default, protected tags honoured, hardcoded exclusions, pre-deletion verification and a timestamped audit log. Ships with a Flask dashboard, PDF and CSV audit reports, and Lambda + EventBridge scheduling. The research behind it was presented at ICANDIT 2026.

Boto3 scikit-learn Flask ReportLab AWS Lambda EventBridge SNS unittest.mock

Seven hand-built levels, nine movement abilities, fourteen trap behaviours, two enemy types — deliberately written as a clean reference codebase rather than a jam prototype. 47 C# files across a four-layer architecture (input, simulation, presentation, persistence), where every playable character is a ScriptableObject, so adding one touches zero lines of code.

Nine UI screens with real confirmation flows, a pooled audio system that never instantiates an AudioSource at runtime, and a Unity CI pipeline that runs edit-mode and play-mode tests and produces Windows and WebGL artifacts on every push. Validated across six build targets.

Unity 2022 LTS C# Cinemachine Unity Input System TextMesh Pro GitHub Actions Git LFS

Languages

ML & Deep Learning

NLP & GenAI

Explainability

Backend & APIs

Frontend & Design

Cloud & DevOps

Data

Game Development

Every repository above follows the same discipline, and I think it's the more interesting half of the work. Packages, not scripts — modules with clear boundaries you can swap or test in isolation. Configuration in YAML with typed dataclasses, not constants buried three files deep. Test suites that run without cloud credentials, a GPU, or a 6 GB download, because a test nobody can run is documentation at best. GitHub Actions on every push. Architecture docs written for someone who isn't me.

And sensible failure modes throughout: the RAG layer falls back to TF-IDF, the recommender falls back to rules, the explanation layer falls back to a template — and the tool that deletes cloud infrastructure won't do it unless you ask twice.

• Winner — Hack The Gap, VIT Chennai (2025). Full-stack platform connecting underserved students to mentors and learning resources; recognised for solution design and practical impact.
• First Runner-Up — Figma × Apple Vision Pro Design Challenge, GDSC VIT Chennai (2023). Spatial-computing UI prototype.
• Paper presented — ICANDIT 2026, INTI International University, Malaysia. Automated AWS Resource Cleanup for Optimization of Cost and Security.
• Outreach Head — TechnoVIT & Vibrance (2024–25). Led a 25-member team across two flagship university festivals.
• Certifications — Google UX Design Professional Certificate · Introduction to Generative AI, Google Cloud (100%) · Python Data Structures, University of Michigan (97.6%)

Away from the keyboard I shoot photography and play badminton. The photography is where most of my design instinct comes from — framing, hierarchy, and knowing what to leave out of the frame turn out to be the same problem as designing an interface.

Open to SDE and AI/ML roles. The fastest way to reach me is email or LinkedIn.