I'm a final-year B.Tech Computer Science and Business Systems student at BMSCE, building practical products at the intersection of software engineering, data science, AI/ML, and Generative AI.
My work focuses on designing reliable end-to-end systems: database schemas, API design, role-based workflows, data and ML pipelines, retrieval systems, and usable full-stack interfaces. I enjoy solving problems where correctness matters — from persistent memory for LLM agents to evidence-first financial reconciliation and retention-focused churn prediction.
I approach engineering with a product mindset: understand the problem, make the design explainable, build for maintainability, and ship something useful.
- Software Engineering roles — SDE, Backend, and Full-Stack
- Applied AI/ML Engineering and Generative AI opportunities
- Open-source collaboration in Python, TypeScript, backend systems, or ML
- Project and research collaboration in Agentic AI, System Design, and Data Science
Python C C++ SQL MATLAB
Pandas NumPy scikit-learn PyTorch CNNs
Feature Engineering Model Evaluation Classification Regression
Exploratory Data Analysis Data Visualization Statistical Analysis
RAG LangChain LangGraph IBM watsonx
Embeddings Vector Search Prompt Engineering
LLM Workflows Persistent Agent Memory Agentic AI
FastAPI Node.js REST APIs Authentication
Role-Based Access Control PostgreSQL pgvector Redis
SQL CSV/API Data Pipelines Database Schema Design
React Next.js TypeScript Tailwind CSS
Responsive UI Development Operator Dashboards Role-Based Workflows
Docker Git GitHub Jupyter Notebook
Google Colab Power BI VS Code Postman
Data Structures & Algorithms OOP DBMS
Operating Systems Computer Networks
System Design Distributed Systems Fundamentals
🧠 Vellum — Cognitive Memory System for LLM Agents
A persistent memory layer for LLM agents designed to support more relevant, self-correcting, and long-context multi-turn interactions.
| Aspect | Details |
|---|---|
| Stack | Python, FastAPI, PostgreSQL, pgvector, Redis, REST APIs |
| Architecture | Episodic memory storage, semantic consolidation, decay-based recall, retrieval, and belief revision |
| Key Engineering | Persistent memory across sessions with context retrieval designed to reduce redundant context retention |
| Impact | Supports more relevant responses and better continuity in multi-turn LLM workflows |
| Repository | View Repository |
Built a cognitive memory system that stores agent experiences, retrieves relevant context through vector search, consolidates information semantically, and supports belief-revision logic when new information conflicts with prior memory.
💰 ReconcileX — Evidence-First AI Finance Reconciliation Controller
A multi-source finance-operations system created for the Razorpay Hackathon. It reconciles synthetic payments, settlements, refunds, and bank credits while ensuring uncertain cases are sent to human review instead of being auto-matched.
| Aspect | Details |
|---|---|
| Stack | Python, FastAPI, PostgreSQL, React, TypeScript, Gemini |
| Data Sources | Payments, settlements, refunds, fees, GST records, and bank credits |
| Matching Policy | Auto-matches only complete and deterministic evidence chains; uncertain cases are routed to human review |
| Safety Controls | Decimal-safe validation, ₹0.01 tolerance, duplicate/idempotent batch ingestion, rule precedence, and deterministic fallback |
| Auditability | Database-enforced append-only audit events and human exception lifecycle controls |
| LLM Scope | Bounded AI assistance for grounded exception explanations and bank-narration reference extraction; no AI financial state changes |
| Validation | 100% precision, recall, and F1 with 0 incorrect auto-matches on a 500-scenario held-out synthetic benchmark |
| Testing | 114 automated tests plus dedicated AI safety and extraction evaluation suites |
| Repository | View Repository |
Designed an evidence-first reconciliation workflow that prioritizes correctness and explainability. The system identifies exceptions such as value variance, delayed settlement, missing references, invalid data, duplicate records, and status conflicts, then presents them through an operator dashboard for drill-down and review.
🎓 Campus Cred — College Activity & Event Management System
A role-based full-stack platform for college event registrations, AICTE activity-point tracking, credential workflows, and administrative operations.
| Aspect | Details |
|---|---|
| Stack | React, Next.js, Tailwind CSS, Node.js, PostgreSQL, Firebase Auth, Docker |
| Users | Students, faculty, proctors, HODs, administrators, and event-management stakeholders |
| Core Features | Event registration, activity-point tracking, administrative workflows, and credential-related management |
| Authentication | Google SSO with role-based access control for secure multi-user access |
| Impact | Reduces manual spreadsheet work by automating registrations and activity-credit tracking |
| Deployment | Docker containerization for portability across local and server environments |
| Repository | View Repository |
Built as an end-to-end college operations platform with separate workflows for different user roles, allowing the college team to manage activities and student participation through a centralized system.
📊 Customer Churn Prediction — Retention-Focused ML Pipeline
An end-to-end customer churn prediction project built on a self-generated synthetic OTT/lifestyle dataset containing 10,000 customer records.
| Aspect | Details |
|---|---|
| Stack | Python, Pandas, NumPy, scikit-learn, Matplotlib, Seaborn |
| Data Pipeline | Data cleaning, EDA, feature engineering, validation, stratified train-test splitting, and target-leakage prevention |
| Models | Logistic Regression and Random Forest, including class-weighted variants |
| Evaluation | Precision, recall, F1-score, macro F1, ROC-AUC, confusion matrices, and threshold analysis |
| Key Results | Logistic Regression achieved 0.6527 ROC-AUC; class-weighted Random Forest achieved 0.6023 macro F1 |
| Retention Outcome | At a 0.30 outreach threshold, identified 711 of 770 churners on the held-out test set — 92.34% recall |
| Business Output | Risk-banded retention outreach exports for medium-, high-, and critical-risk customer cohorts |
| Repository | View Repository |
Converted model predictions into actionable retention segments rather than treating churn prediction as only a classification task. The project uses threshold analysis to balance outreach cost with churner coverage.
🚦 AI-Based Traffic Congestion Monitoring & Alert System
A multi-agent AI prototype built for the IBM Agentic AI Hackathon to analyze traffic data, detect congestion patterns, predict hotspots, and generate proactive route advisories.
| Aspect | Details |
|---|---|
| Stack | Python, IBM watsonx Granite, LangFlow, LSTM, ARIMA, Prophet |
| Architecture | Specialized agents for data analysis, congestion trend detection, and alert/advisory generation |
| Workflow | Traffic data ingestion → trend and anomaly detection → short-term forecasting → natural-language advisories |
| Forecasting Goal | Short-term congestion prediction for the next 2–4 hours |
| Impact | Supports proactive traffic management through early alerts and route recommendations |
| Repository | View Repository |
Designed a multi-agent workflow in which each agent handles a specific responsibility: analyzing incoming traffic signals, identifying congestion trends and likely hotspots, and generating understandable alerts for users or operations teams.
Hemvat Consultancy Services Private Limited · July 2025 – September 2025
Worked on end-to-end data science tasks for consulting-driven datasets, with a focus on converting raw data into decision-ready insights.
- Performed exploratory data analysis to identify trends, outliers, and data-quality issues
- Engineered features for downstream analytics and model readiness
- Built data visualizations to communicate findings clearly to stakeholders
- Documented data pipelines and analysis workflows for reproducibility
Python Pandas EDA Feature Engineering Data Visualization
| Category | Focus Areas |
|---|---|
| Building | Reliable full-stack products, AI-agent memory systems, finance-operations workflows, and ML pipelines |
| Learning | Advanced system design, distributed systems, LLM orchestration, agentic architectures, and cloud deployment |
| Exploring | Retrieval systems, vector databases, model deployment, feature stores, and AI safety for tool-using agents |
| Preparing For | SDE, Backend, Full-Stack, AI/ML Engineering, and applied Generative AI opportunities |
"Engineering is not about writing code — it's about designing systems that scale, adapt, and last."