CVE Twin
Repair a vulnerability without breaking the services that depend on it.
Read case studyWatch walkthroughINDUJ GUPTA FULL-STACK ENGINEERBENGALURU, INDIA / 2026
AI PRODUCTS. SECURITY SYSTEMS.Intelligence.
Computer Science, MIT Bengaluru
Class of 2027
Intelligence.
Engineered.
I build across the stack.
Then look for the places it can break.
Computer Science, MIT BengaluruClass of 2027
FIG. 01 / CVE TWINDEPENDENCIES → REPAIR → EVIDENCE
INSIDE THE DEPENDENCY SYSTEM DEPENDENCY MODEL
Every connection has a consequence.
A spatial interpretation of CVE Twin: one vulnerable dependency, the services around it, and the evidence needed to repair it.
Spatial illustration of a controlled local demo. The hosted explorer does not execute repairs.
SELECTED WORK / 01—04Four systems.
Four systems.
Built to be inspected.
Real interfaces. Open implementation.
The problem, the decisions, and the evidence behind each build.
- Actual interfaceSECURITY / AGENTIC AI
- Actual interfaceAPPLIED ML / FINTECH
Fraud Risk Manager
Turn fraud scores into explainable decisions within a review budget.
Read case studyWatch walkthrough - Actual interface · demo modeMULTI-AGENT SYSTEMS
SentinelMesh
Bring crisis signals together before a human makes the call.
Read project - Original experimental figureCOMPUTER VISION / RESEARCH
TCVM-Net
Check whether traffic-vision predictions stay consistent through interruptions.
Read project
Security engineering / recorded local repair
CVE Twin
Follow a repair from a failing baseline to a change that passes the same checks.
Recording transcript
- The baseline shows 4 tests passing and 3 failing.
- Codex activity logs show diagnosis, repair, and verification.
- The final result shows 7 tests passing and 0 failing.
This silent clip records one local demonstration. The three-strategy comparison is documented separately in the case study.
Follow the evidence
Original project capture01 / 03
Step 1 of 3
Read the recorded runStart with tests that fail.
Before changing code, the local workflow demonstrates the vulnerability. The recorded baseline has four passing tests and three failing security checks.
- Passing before
- 4
- Failing before
- 3
Read this image
The left-hand baseline panel identifies the three failing checks. A suite that was already green would not establish a repair.
Recorded demonstration result7 / 7 tests passing
A controlled fixture models the Log4Shell vulnerability class. All three recorded repair candidates were viable; the smallest change was selected. The hosted explorer cannot run the local repair engine.
Read the full caseRead all three steps
Start with tests that fail.
Before changing code, the local workflow demonstrates the vulnerability. The recorded baseline has four passing tests and three failing security checks.
Recorded result / read the BEFORE CODEX panelThree strategies. Shared checks.
Minimal change, parser hardening, and a provider allow-list run in isolated worktrees. Each faces the exploit, security tests, and dependent-service contracts. All three were viable in the recorded comparison.
Actual investigation interface / comparison results are in the linked runThe smallest viable change wins.
The recorded selection is the minimal repair. The security suite finishes with seven passes and no failures, while the result keeps the diff, branch, commit, and verification evidence available for review.
Recorded result / read AFTER CODEX and the proposed change
Applied machine learning / decision walkthrough
Fraud Risk Manager
See how model assessments become an explainable decision—and when that decision needs a person.
Follow the evidence
Original project capture01 / 03
Step 1 of 3
Read the model architectureTwo models inspect the transaction.
XGBoost looks for learned fraud patterns; an Isolation Forest verifier assesses unusual structure. The captured example shows detector and verifier scores alongside the features contributing to the decision.
- Learned fraud patterns
- Detector
- Unusual structure
- Verifier
Read this image
Read DUAL-MODEL EVALUATION, then the SHAP contributions below it. These scores describe this captured example; they are not calibrated probabilities.
Separate offline detector evaluation40.6% lower modeled loss
The 40.6% result compares the capacity-capped detector with an amount-threshold baseline on an IEEE-CIS temporal holdout. It uses modeled costs, not production savings. The review branch here illustrates the policy; it is not a new scored transaction.
Read the full caseRead all three steps
Two models inspect the transaction.
XGBoost looks for learned fraud patterns; an Isolation Forest verifier assesses unusual structure. The captured example shows detector and verifier scores alongside the features contributing to the decision.
Actual interface capture / transaction #900006178 was blockedAgreement permits an action.
Both high-risk assessments produce a block; both low-risk assessments permit an allow. The actual trace shown here records agreement on high risk, followed by BLOCK and an audit-log entry.
Actual decision trace / the recorded action is BLOCKDisagreement goes to review.
If the assessments disagree, the policy holds the transaction for a human instead of choosing one model. This step illustrates that branch using the real operations interface; it is not a replay of the blocked transaction above.
Actual operations interface / disagreement branch is illustrative
CONTINUING THE WORK
Two more systems.
Follow the agent workflow or the temporal-vision pipeline, then inspect the original evidence.
2026
MULTI-AGENT SYSTEMSSentinelMesh
Solo build / full-stack engineering & agent orchestrationA crisis-intelligence command center connecting maritime, cyber, market, weather, and news signals. Specialist agents assemble evidence while consequential decisions stay behind human approval.
Read projectWorkflow guide
Gather evidence
Collect the available public maritime, market, weather, news, and cyber signals, with fallbacks for unavailable sources.
Documented agent workflow · The screenshot shows the interface in demo mode.
2026
COMPUTER VISION / RESEARCHTCVM-Net
Solo build / computer vision & experimental evaluationTemporal verification for traffic vision: confidence stability, motion continuity, and short-gap recovery. The public experiments examine behavior under controlled perturbations, including a 20-clip occlusion stress test.
Read projectWorkflow guide
Prepare separate video clips
Split by source clip and generate perturbations after the split, keeping calibration and test sequences separate.
The figure above comes from the separate 20-clip occlusion study. This guide describes the documented implementation and evaluation workflow.
Experience / 2024—2026The work
The work
within teams.
Product engineering, security, and the systems behind an interface.
Read the résuméProduct engineeringJun 2026 — 1 Aug 2026
Turbostart
Software Engineering Intern
Bengaluru, IndiaMy contribution
Contributed to product development at Turbostart as a Software Engineering Intern, collaborating with the engineering team.
- Product development
- Engineering collaboration
Software engineering internshipCompleted 1 August 2026
Full-stack developmentStarted Dec 2025
Prizlyn Technologies
Software Engineering & AI Development
Bengaluru, IndiaIncluding Confab 360 Degree and MentorEdge
My contribution
Full-stack and AI development across Prizlyn, Confab, and MentorEdge. Developed React, TypeScript, and Node.js features across three major product releases and built GitHub Actions pipelines that reduced build-and-deploy time by 30%.
- React
- TypeScript
- Node.js
- GitHub Actions
Résumé-reported outcomes
- 3product releases
- 30%faster build & deploy
Applied securityJul 2025 — Aug 2025
UP Police
Cybersecurity Intern
Lucknow, IndiaMy contribution
Engineered AI helmet and seatbelt detection with PyTorch and OpenCV. Conducted penetration testing with Burp Suite, Nmap, and Metasploit, identifying 12+ critical vulnerabilities and reducing incident response time by 60%.
- PyTorch
- OpenCV
- Burp Suite
- Nmap
- Metasploit
Résumé-reported outcomes
- 12+critical vulnerabilities identified
- 60%less incident response time
APIs and ML pipelinesJun 2025 — Aug 2025
Karma Labs
Software Development Engineer Intern
RemoteMy contribution
Built RESTful APIs and React and TypeScript frontend components. Asynchronous processing optimizations reduced ML pipeline latency by 35% and increased user engagement by 25%.
- REST APIs
- React
- TypeScript
- Async processing
Résumé-reported outcomes
- 35%lower ML pipeline latency
- 25%higher engagement
APIs and data accessJun 2024 — Aug 2025
TruckHai
Full Stack Engineer · Contract
RemoteMy contribution
Owned REST API development for 500+ enterprise clients. Implemented JWT and OAuth 2.0 and optimized PostgreSQL and MongoDB queries for 40% faster data retrieval.
- JWT
- OAuth 2.0
- PostgreSQL
- MongoDB
Résumé-reported outcomes
- 500+enterprise clients
- 40%faster data retrieval
Career outcomes from my résumé. Explore project evidence
RESEARCH / FOUR PAPERS03
Research
& questions.
Exploring the tradeoffs between intelligence, explainability and the resources a system actually has.
Multi-Objective Edge Resource Optimization with D3QN-PER in Quantum-Classical IoT Edge Networks
Abstract & scope
Can reinforcement learning schedule bursty quantum-key-distribution workloads without overloading edge nodes? Co-authored a D3QN-PER study evaluated in a three-node simulation, with five-seed validation against Round-Robin and Least-Loaded baselines. Published at IEEE COMSNETS; the evidence concerns simulated scheduling workloads.
Learning Quantum Measurement Probabilities with Transformer Networks using Qiskit Simulations
Abstract & scope
Can transformer networks learn the probabilities of quantum measurement outcomes? Explores that question using Qiskit-generated simulations. The work concerns learning from simulated quantum systems; it does not establish performance on physical quantum hardware.
An Explainable and Resource-Efficient Transformer Pipeline for CPU-Based Document Summarisation and Question Answering
Abstract & scope
How can document summarisation and question answering run on CPU-only machines while remaining explainable? Co-authored a transformer pipeline focused on resource-constrained deployment. Accepted at AICCoNS 2026; the linked co-author announcement establishes the scope and acceptance, without reporting a deployment benchmark.
Bridging the Accuracy-Sustainability-Explainability Trilemma in Air Quality Informatics: A Carbon-Aware Consensus Framework
Abstract & scope
How should an air-quality model balance prediction quality, computational cost, and an explanation people can inspect? Co-authored the Carbon-Aware Consensus framework to study those tradeoffs. Accepted at IEEE CONECCT 2026; the linked announcement does not establish a deployed monitoring system.
A LITTLE CONTEXT04
A little
about me.
FULL-STACK ENGINEERING / AI & SECURITY
I build across the stack, and think about the whole system.
Full-stack engineering for AI and security products, from API design and model integration to interfaces that make complex decisions inspectable. My work spans developer tooling, applied machine learning, product interfaces, and the boundaries that keep a system secure.
I’m studying Computer Science at MIT Bengaluru, graduating in 2027. Since August 2025, I have led SDG Club: a 50-member team whose initiatives have reached 500+ students, creating opportunities to build, share, and learn.
The complete résuméBuild the product
Interfaces, APIs, and the systems connecting them.
Secure the system
Authentication, threat modeling, and hands-on security.
Explore intelligence
Applied computer vision and quantum computing research.
Ship the infrastructure
Cloud tooling, delivery pipelines, and Web3.
EDUCATION / 2023 — 2027
B.Tech in Computer Science
Manipal Institute of Technology, Bengaluru
Data Structures & Algorithms · Computer Networks · Operating Systems · Cryptography · Machine Learning · Distributed Systems
CREDENTIAL / October 2024
Ordered Data Structures
University of Illinois Urbana-Champaign
Credential 40K83BGOCR8ZRECOGNITION06First-place finishes.
06First-place finishes.
Built with a team.
Hackathons are where I test ideas under pressure, learn from collaborators, and turn a problem statement into something working.
COMMUNITY
Beyond the build.
Organizing, sharing what I learn,
and representing the communities I’m part of.
CYBERSECURITY / COMMUNITY
CryptNite 2026
Helped organize CryptNite at MIT Bengaluru, building and deploying CTF infrastructure for national qualifiers and on-campus finals.
- Teams
- 250+
- Participants
- ~1,000
FEBRUARY 2026 / MAHE BENGALURU
India AI Impact Summit
Selected to represent MAHE Bengaluru at the OpenAI Town Hall at IIT Delhi and the India AI Impact Summit 2026 in New Delhi.
Read the story
SEPTEMBER 2025 / SDG CLUB PRESIDENT
HackSpace ’25
As President of SDG Club at MIT Bengaluru, I led our flagship hackathon with Honeywell Aerospace Technologies. Our core committee and volunteers brought 70 shortlisted teams together to tackle industry problem statements.
- Hackathon
- 48h
- Registrations
- 250+
- Shortlisted teams
- 70
OCTOBER 2025 / SDG CLUBSpeaking at the SDG Innovation Sprint
Spoke alongside Panchadip Bhattacharya about research publications, launching startups, and hackathon strategies. The two-day SDG Club initiative received 250+ registrations and brought teams together to ideate, prototype, and pitch.
Read the club recapDEVELOPER COMMUNITY / TEAM PRESENTATIONSonar / DevRelSquad developer conference
Presented our work with Saaheer Purav, Ashmika Jain, and Riddhi Rajesh at a developer conference hosted by Sonar and DevRelSquad (by GoAvo.ai), at the Microsoft office. A chance to exchange ideas about AI-assisted development with fellow engineers.
Read my teammate’s accountOFF THE CLOCK / INTERACTIVE STUDIES
A place to play.
Three small experiments in connection, interference, and signal. Change a parameter and see what happens.