This is a long virtual resume that contains everything I did since I chose Software Engineering, Applied Research and overall Product Building as a career for the time being. Might become a goose farmer later in life, who knows :)
Work Experience
Member of Technical Staff
Sqwish AI
Currently building low-latency contextual bandit inference for AI Agents
Open Source Maintainer
Jenkins
Maintaining the Jenkins GitLab Plugin used by over 60,000+ developers worldwide
- Migrated the Jenkins GitLab Plugin from RESTEasy Library to GitLab4J-API with Reverse Proxy support
- Adapted the Docker Maven Plugin for GitLab4J-API library
- Improved the Docker-based test suite by Migrating 500+ Unit and Integration tests and improved overall code coverage
as a Google Summer of Code mentee
Mentoring Google Summer of Code contributors in Jenkins since 2024 for the development of a vertical self-learning multi-agent AI diagnostic system to automate failure diagnosis for over 2000+ plugins and complex CI/CD pipelines
Skills involved - Java, Python, Ngnix, Docker, RAG, MCP and A2A, Langchain, Qdrant, Pydantic, list just goes on...
Protocol Enginnering Grantee
Ethereum Foundation
Implemented the Fast Confirmation Rule in Lighthouse Consensus Client with EF Protocol Consensus Team
Software Engineer Intern - Linux Foundation
Cloud-Native Computing Foundation
- Designed and implemented mTLS via HBONE tunneling for metrics endpoints in Istio's Ambient Mesh aligning with zero-trust principle. Became a Member of Istio's Networking Working Group upon the completion of the internship work.
- Implemented admin operations security controls in Thanos UI that restricts critical admin-level operations at both API and UI levels, improving production security posture for large-scale monitoring deployments - PR
- Added comprehensive network security policies for Kubernetes-deployed Thanos components using NetworkPolicy resources to establish network isolation and reduce attack surface in multi-tenant environments
- Developed dynamic, color-coded query result visualizations in the Thanos Query UI to enhance operational efficiency for monitoring teams analyzing distributed Prometheus time-series data.
Skills involved - Go, Rust Typescript, React.js, Docker, Kubernetes, Helm, Prometheus, Grafana
Check out my blog on contributing to Istio to learn more.
Open Source Contributor
uv
- Added CLI flag to display compressed package sizes aiding dependency analysis - PR
- Added support for GitLab as a trusted publisher, implementing secure OIDC token discovery to enable passwordless authentication with PyPI - PR
- Added authoritative constraint of local pin over global in the UV CLI, prioritizing project-scoped .python-version files over global user preferences when conflicts arise - PR
- Added a context-aware warning in CLI to help with conflicting VIRTUALENV - PR
Skills involved - Rust, Python
Software Engineer Intern - Summer of Bitcoin
Fedi
- Designed and Implemented the Escrow Module for the Fedimint ecosystem based on trustless and dispute-resistant private escrow scheme
- Upgraded the module template to support latest long term stable Fedimint protocol, while maintaining the backward compatibility, for easing the process of module creation for the Fedimint ecosystem
- Refactored the codebase for reducing redundant dependencies and improving maintainability and added some UI features.
Checkout my journey of contributing to Fedimint in my blog - Spent my Summers with Fedimint!
Skills involved - Rust, TypeScript, Docker, Prometheus, OpenTelemetry, WASM, Bitcoin, Lightening Network
Undergraduate Research and Projects :
Undergraduate Research
Advisors: Prof. Dibakar Ghosal and Subhajit Roy (Department of Earth Science & Computer Science)
Modelling of Forward and Inverse Waveform Inversions via Finite Basis Physics-Informed Neural Networks for Seismic Data Interpretation on an actual Supercomputer!
- Architected multiscale Finite Basis Physics Informed Neural Network for subsurface modelling, solving 2D acoustic and elastic wave equations using JAX’s JIT compilation for 2x performance improvements in neural network training, implementing domain decomposition with partition-of-unity windows, optimized architectures for tackling spectral bias during full waveform inversion
- Designed scalable model-parallel architecture for arbitrary subdomain distribution across GPU clusters, implementing topology-aware sharding with NVLink field exchange (300GB/s), block-accumulation synchronization, and chunked evaluation enabling memory-efficient training. Demonstrated on V100s with 75 subdomains achieving 15x speedup and 50% per-device memory reduction, with architecture supporting 1000+ subdomain scaling on multi-node HPC systems
- Engineered memory-efficient evaluation pipeline for V100 HPC clusters through chunked inference eliminating OOM errors, hybrid 4th/2nd-order CPML stencils for PML stability, and adaptive spatial oversampling (6x) with nearest-neighbor downsampling preserving sharp wavefronts
- Integrated Fourier Feature Embeddings and SIREN architectures with per-subdomain frequency tuning, allowing the model to capture high-wavenumber components 10x faster than standard MLPs.
- Integrated Extreme Learning Machines with FBPINNs, replacing traditional gradient descent with Rank-Revealing QR (RRQR) linear solves, achieving 10-100x reductions in training time while maintaining numerical stability.
- Architected a multi-stage training strategy (Adam → L-BFGS → Damped Newton) to navigate highly non-convex loss landscapes, reducing final L2 error by up to 14x compared to first-order optimizers.
- Developed RAR (Residual-based Adaptive Refinement) and SAAR (Self-Adaptive Residual Sampling) to dynamically concentrate collocation points in high-residual regions (e.g., salt-dome boundaries), improving spatial resolution by 2x without increasing total point budget.
- Implemented derivative-penalized loss functions to enforce C¹ regularity across overlapping subdomain interfaces, significantly reducing Gibbs-like artifacts in sharp-front wave propagation
- Built a comprehensive visualization suite for Full-Waveform Inversion, including real-time loss tracking, seismological wiggle plots, shot gathers, and spatial/temporal error analysis for ground-truth validation against CPML solvers.
Skills - Python, Fortran, SLURM GPUs (Param Sangranak), numPy and cuPy, JAX, Linux, Bash
Undergraduate Project
Advisor: Prof. Yatinder Nath Singh (Department of Electrical Engineering)
An asynchronous multidimensional in-memory distributed hash table implementation based on Chord - Pikachu
- Engineered an asynchronous, in-memory distributed hash table with multi-dimensional keyspace support with an efficient Chord overlay network, including finger tables and recursive lookup logic, achieving average O(log N) hop count for key resolution
- Built a resilient routing maintenance subsystem with periodic stabilization, successor/predecessor checks, and finger-table repairs to ensure network consistency under churn
- Engineered connection-pooled gRPC layer with backpressure handling and bounded buffers, reducing connection churn and ensuring stable performance during high-volume key transfers
- Implemented robust server lifecycle management with readiness signalling, bounded startup timeouts, and graceful shutdown for clean drain of in-flight requests
- Introduced streaming handoff APIs for efficient ownership transfer of large key ranges during join/leave, reducing tail latencies and avoiding head-of-line blocking
Skills - Rust, gRPC, libp2p, Distributed Systems, Consistent Hashing
Undergraduate Project
Advisor: Prof. Adithya Vadapalli (Department of Computer Science)
A privacy-preserving discovery service. - Rumi
- Engineered a privacy-preserving contact discovery service using double-sided blinding over elliptic curves, a Path Obilivous RAM backend store, and zero-knowledge set-membership proofs to ensure the server learns nothing about queries, matches, or access patterns
- Built end-to-end blinding protocol using domain-separated hash-to-curve and compressed point encodings
- Designed prefix-based bucket selection paired with fixed dummy accesses and constant response shaping on top of Path ORAM for efficient server-side filtering while maintaining indistinguishability; tunable privacy/performance balance
- Benchmarked ORAM performance across read/write, sequential/random patterns, and variable payload sizes
- Hardened sensitive data handling with explicit secure wiping of ORAM blocks, stash, and internal buffers; minimized side-channels via fixed-size responses and uniform access paths
Skills - Rust, Stream Processing, Cryptography