AVAILABLE — SPRING / SUMMER 2027NEW YORK, USA

I build softwaresystems that hold upin production.

Software engineer and MS Computer Science candidate at Stony Brook, graduating May 2027. I work where correctness is measurable — consensus protocols, surrogate models, and inference that has to run on one machine, alongside production web platforms that ship.

CURRENT STATE
degree
MS CS, Stony Brook
gpa
3.61 / 4.00
graduating
May 2027
published
APS GEC 2025
C++·RAFT & MULTI-PAXOS·CUDA·PYTORCH·DISTRIBUTED SYSTEMS·PYTHON·LANGGRAPH·NEXT.JS·FASTAPI·CONCURRENCY·ON-DEVICE INFERENCE·
45%
FASTER PIC PLASMA
SIMULATION
200+
TPS UNDER 3-REPLICA
CONSENSUS
5
MODELS ORCHESTRATED
ON ONE LOCAL GPU
1st
BAND OF AGENTS
HACKATHON 2026
01 — SELECTED WORK

Eight systems, eight measurable results.

click a row to expand ↓

Built production features for a B2B industrial sourcing platform, architecting a serverless backend on Firestore, Resend and HubSpot CRM for data persistence, transactional email and automated lead synchronization. Developed an interactive landed-cost calculator modeling freight, duty and handling, alongside 15+ responsive pages spanning product, catalog, industry and RFQ workflows. Improved production readiness through site-wide WCAG 2.1 AA accessibility, security, SEO and performance work, including JSON-LD structured data and deployment hardening.

15+ RESPONSIVE PAGES · WCAG 2.1 AANEXT.JS · REACT · TYPESCRIPT · FIRESTORE · RESEND · HUBSPOTSOURCE ↗ for Aaron Technologies
The Aaron Technologies homepage: an industrial sourcing site headlined "Your BOM, one partner, 30+ factories".

A distributed supply-chain platform that detects vessel delays from a live AIS feed and dynamically recruits seven specialized agents across three frameworks via Band to coordinate logistics incident response. Auditable human-in-the-loop workflow — quorum voting, adversarial dissent review, and demurrage estimation grounded in published carrier tariffs — emitting traceable decision dossiers from full room history.

7 AGENTS / 3 FRAMEWORKSLANGGRAPH · PYDANTICAI · CREWAI · BANDSOURCE ↗ for FreightRoom
The FreightRoom incident console: a five-phase stepper from detection to approval gate, a live AIS status pill, and an accruing demurrage cost card.

Implemented Raft consensus — leader election, replicated logs, fault-tolerant recovery — for a distributed key-value system, validated under simulated node failures and network partitions. Extended it with a deterministic-ordering transaction engine using a global sequencer and Multi-Paxos to guarantee strictly serializable execution.

200+ TPS / 3 REPLICAS390+ TPS SINGLE-REPLICA · TPC-C ON MAKO · MASSTREESOURCE ↗ for Raft Consensus & Multi-Paxos Sequencer
Architecture: a client transaction enters partition 0, where a deterministic scheduler submits a slot to the Multi-Paxos coordinator, replicates to two Paxos replicas, executes against Masstree, and broadcast-dispatches to a worker partition.

A fully on-device, privacy-preserving multimodal AI platform orchestrating five models (Cosmos-Reason2-8B VLM, CLIP ViT-L/14, Faster R-CNN ResNet50-FPN, Qwen3 embedding, Nemotron reranking) to ground construction-site video findings in OSHA and NYC DOB regulations via retrieval-augmented search.

5 MODELS, ONE LOCAL GPU7 LIVE NYC OPEN DATA SETSSOURCE ↗ for Argus — Local AI Compliance Platform
A New York City sidewalk excavation — a source video frame of the kind Argus grounds against OSHA and DOB rules.

A multi-agent middleware that applies anonymous peer review to prompt engineering, coordinating rewriter, reviewer, and chairman agents via LangGraph. Validated in a 13-user study, with an ablation study isolating the effect of reviewer-model diversity.

4.77/5 MEAN TRUST, 0% OVERRIDE+25.5% CONSENSUS STRENGTHSOURCE ↗ for ConsensusPrompt
Ablation chart: consensus strength rising from 63% with three rewriter models to 79% with five.

An AI-powered developer onboarding platform with a custom MCP server and FastAPI backend that automates understanding of unfamiliar, large software systems. Real-time WebSocket workflows and repository retrieval surface dependency graphs, entry points, code hotspots, and project conventions.

AUTOMATED REPO COMPREHENSIONFASTAPI · MCP · GITPYTHONSOURCE ↗ for OnboardOps
OnboardOps architecture: a FastAPI backend exposing seven Git analysis tools over MCP, bridged to a Next.js dashboard by WebSocket events.

A production platform converting cooking videos into typed, scalable recipes — a Next.js BFF speaking gRPC to a Python extraction service, over a Redis job queue with idempotency, exponential-backoff retry, dead-lettering and caching. A dependency-free TypeScript engine applies cooking semantics rather than arithmetic: seasoning grows sublinearly, countable ingredients round to whole units, baking ratios stay exact — verified by property-based tests held at 100% branch coverage in CI. When a creator's description carries no recipe, sampling put that at one in five, the pipeline falls back to reading the video itself. Provisioned in Terraform, deployed to Cloud Run by keyless OIDC federation.

657 TESTS · 100% BRANCH COVERAGENEXT.JS · TYPESCRIPT · PYTHON · GRPC · REDIS · POSTGRES · TERRAFORM · GEMINILIVE ↗ site for Mise — Video-to-Recipe Extraction PlatformSOURCE ↗ for Mise — Video-to-Recipe Extraction Platform

Intercepts the clipboard payload the moment company data is about to leave a laptop for an unapproved cloud AI tool, inspects it locally, blocks it with a cited policy clause, and re-routes the question to a 35B model on the box so the employee still gets an answer. Nothing leaves the machine: every verdict reports bytes_egressed: 0, and the agent runtime is denied egress at the proxy — enforced outside the sandbox, unmodifiable from within it. Measured at a 4.00% false-positive rate over 1,000 benign pastes drawn from six independently-licensed sources, with zero false positives across 480 items of Stack Exchange, Wikipedia, MBPP and HumanEval.

4.00% FPR · 0 BYTES EGRESSEDCHROME MV3 · PYTHON · vLLM · MONGODB · NEMOTRONSOURCE ↗ for Airlock
An Airlock verdict: a customer-record paste blocked before leaving the machine, citing policy POL-004, with classifier confidence 0.94, a 312 ms decision, and bytes egressed 0.
EARLIER WORK
UNDERGRADUATE · 2021—2024
MOBILE & FULL-STACK
Canteen Automation System
Led an agile team building a mobile-first canteen ordering and inventory platform for campus dining.
+30% ORDERING EFFICIENCY
Flutter · Agile · Team lead
NETWORKING & SYSTEMS
Local File Sharing System
Peer-to-peer transfer over IPC and low-level sockets, engineered for zero-loss delivery on a local network.
OPTIMIZED P2P THROUGHPUT
C++ · IPC · Sockets
AI SECURITY & VISION
MagNet — Adversarial Defense
A two-pronged detector and reformer defense against adversarial attacks on deep image classifiers.
HIGH DETECTION RATE ON NOISE
PyTorch · Computer vision
DATA STRUCTURES
Vaccine Record Management
Custom C++ data structures and persistent file handling for high-speed record lookup and integrity checks.
CONSTANT-TIME LOOKUP
C++ · File I/O · Memory
02 — EXPERIENCE

Research experience.

JAN 2024 —
JUL 2024
Integrating DL with 2D-3V Particle-in-Cell Simulations of Low Temperature Plasmas
DHIRUBHAI AMBANI UNIVERSITY · PRESENTED AT APS GEC 2025
  • Replaced the charge-deposition, Poisson-solver and electric-field stages with a U-Net surrogate model, achieving a 45% execution-time reduction on 128×128 grids with 2.6e5 particles.
  • Enforced physical fidelity with Physics-Informed Neural Networks, maintaining mean percentage error under 5% against baseline PIC-MCC across 500-iteration E×B test cases; implemented in Python with C extensions on Unix/Linux.
MAY 2023 —
AUG 2023
Summer Research Intern — Software Testing & Test Case Prioritization
DHIRUBHAI AMBANI UNIVERSITY, GANDHINAGAR, INDIA
  • Benchmarked four test-case prioritization techniques across 30 open-source Java projects up to 87K SLOC and 5,000+ test cases, using runtime and mutation-based APFD.
  • Profiled execution bottlenecks and automated evaluation pipelines in Python, cutting pipeline runtime by 30%.
03 — EDUCATION

Foundation.

AUG 2025 — MAY 2027 (EXPECTED)
MS Computer Science
Stony Brook University, Stony Brook, NY
GPA 3.61 / 4.00Distributed SystemsMachine LearningSoftware Engineering
NOV 2020 — MAY 2024
B.Tech. (Honours), ICT
Dhirubhai Ambani University (formerly DA-IICT), Gandhinagar, India
GPA 3.45 / 4.00Minor: Computational ScienceHonours
04 — STACK

Sorted by depth, not by count.

DEEP — DAILY
C++PythonRaft & Multi-PaxosConcurrencyCUDAPyTorch
SHIPPING PRODUCTION
Next.jsTypeScriptReactFastAPILangGraphFirestoreDockerCI/CD
WORKING KNOWLEDGE
Java · Kubernetes · AWS · Azure · GCP · PostgreSQL · Redis · MongoDB · MCP · FAISS · scikit-learn · MATLAB
05 — CONTACT

Full-time software, ML and systems roles from Spring 2027.

Open to roles, collaborations, or a technical conversation about consensus, surrogate models, or on-device inference.