Computer Science (Cyber Security) undergraduate — building at the intersection of
machine learning, full-stack engineering, and security.

• B.E. in Computer Science & Engineering (Cyber Security) at MS Ramaiah Institute of Technology — CGPA 9.52 / 10.
• Project Intern at Samsung R&D Institute India (PRISM) — building and compressing a multimodal vision-language model for on-device emotional and contextual captioning.
• Co-author of a Springer Nature book chapter on Quantum Computing for IoT.
• I like taking hard problems end to end — ML systems, production web apps, and security tooling — with tests, CI, and measurable results.
• LeetCode Knight, max contest rating 1858 (top 6.13% globally, 22 rated contests).

Reconciles a merchant's order ledger against a payment gateway's report and the bank statement — three systems recording the same money that never agree.

• Built a data-processing engine that reconciles a merchant order ledger, a payment-gateway report, and a bank statement through tiered deterministic matchers, resolving 90.8% of 141 entities at 100% accuracy across 14 defect classes. Every unresolved record lands in a categorised exception report with its reason.
• Implemented the O(n³) Hungarian algorithm from scratch to resolve contested bank-to-settlement matches jointly rather than greedily, behind a verification gate encoding paise-exact money, MDR, and T+1 settlement rules.
• Sustains ~233K entities/sec, flat from 141 to 5,022 entities, and is guarded by 111 tests whose adequacy was verified by mutation testing plus a CI pipeline that asserts the measured accuracy, so a regression fails the build.
• Scope stated honestly: the data is synthetic and generated by the same author, so the numbers show the engine catches the defect classes it was designed around, not performance on a real merchant's books.
• Python · FastAPI · Uvicorn · pytest · mutation testing · GitHub Actions · Render
• Live dashboard · Live app · Five-minute walkthrough

An LLM proposes vulnerabilities; Sentinel counts one only after an exploit has run against the real code in a network-off, unprivileged Docker sandbox and a line tracer confirms the accused line executed. Every candidate is graded on a five-tier evidence ladder, and each fix is checked by replaying the exploit that proved the bug.

• Strict recall 60–80% and strict precision 75–100% over three runs on a five-bug benchmark. Two runs had a false proof on the clean control; it is published with the results and is the top open issue.
• A self-review found and fixed a critical bypass (an exploit could grade itself), plus gate, report-XSS and prompt-injection issues. Every fix is pinned by a regression test that fails when the fix is reverted.
• 340 offline + 15 Docker tests; SARIF 2.1.0 output and --fail-on for CI; any model through LiteLLM (developed on Ollama, Groq, Gemini).
• Python · LiteLLM · Docker · AST taint analysis · pytest · GitHub Actions
• Project page · Sample report

Turns one sentence into a structured course, streamed lesson by lesson, with quizzes, flashcards, a per-lesson tutor and scored mock interviews. Most of the code is about what happens when the model is slow, wrong or down.

• Engineered a custom multi-provider AI router (Gemini → Groq → OpenRouter) with per-provider circuit breakers, retry with backoff, and timeouts, so an upstream outage or rate limit degrades the response instead of failing it.
• Cut perceived latency by streaming courses lesson-by-lesson over SSE, so users see first output in seconds instead of waiting on a multi-minute generation. Secured it with rotating refresh tokens and per-route rate limiting.
• Built the quality gate alongside the feature: unit and integration tests, Playwright end-to-end suites, and autocannon load tests in CI, plus an LLM-as-judge eval harness that fails the build on output-quality regressions.
• React · TypeScript · Node.js · Express · MongoDB · SSE · Zod · Jest · Playwright · autocannon · GitHub Actions
• Live demo

Scan a real cube with your webcam, verify the colours, and watch a hand-written Kociemba two-phase solver drive an animated 3D solution.

• The solver is written from scratch (cubie model, coordinate reduction, pruning tables, IDA* search) and runs in a Web Worker: ~20.6 moves on average and correct on 1,500+ random cubes, checked against an independent engine.
• Webcam scanning with HSV colour classification, animated Three.js playback, and an installable offline PWA. 66 unit tests plus Playwright end-to-end tests.
• TypeScript · Three.js · Vite · Playwright
• Live demo

Detects DGA domains, DNS tunnelling and exfiltration, from sensor to analyst dashboard, with SOAR-style containment and a human in the loop.

• Built a real-time detection platform: a FastAPI service over SQLAlchemy with 21 REST endpoints and SSE live streaming to a React dashboard. Availability is handled explicitly: a liveness probe returns 503 when the database degrades, and containment rules auto-expire after 24 h so a false positive cannot wedge the network.
• Traced 407 false positives to the default 0.5 decision threshold, not the features, and recalibrated against an explicit false-positive budget: 407 → 21 at unchanged 100% recall on 12,000 domains across 4 DGA families.
• Python · FastAPI · SQLAlchemy · SQLite · React · scikit-learn · Docker Compose · pytest · CodeQL
• Live demo

A constrained model predictive controller that drives the production choke on a naturally flowing oil well to a requested oil rate — and when that rate is not safely reachable, says so, names the limit in the way, and produces the most it can instead. Written for Honeywell Campus Connect (hackathon round 2).

• 0 constraint violations across 30 nominal runs; under randomised model error in the hardest scenario, 7 of 150 runs brushed a limit (worst: 1.16 psi on a 2850 psi limit).
• +5.6% oil over a cautious operator at the same zero violations; a conventional PI controller violates on 69% of intervals. Control model R² 0.983–0.993 on Honeywell's independent reference data.
• Python · MPC · Streamlit
• Try it in the browser

• EmoCapNet (Samsung PRISM) — multimodal vision-language model for emotion-aware captioning, compressed for on-device use. Currently writing it up for ICASSP.
• Springer Nature book chapter — Quantum Computing for IoT.

Languages — C++, Python, TypeScript, JavaScript, SQL
ML — PyTorch, scikit-learn, Hugging Face Transformers, SHAP
Web — React, Next.js, FastAPI, Node.js
Infra & tooling — Docker, GitHub Actions, pytest, Jest, Linux

Foundations — Data Structures & Algorithms (C++), Operating Systems, DBMS, Computer Networks, OOP

Open to software engineering internships. The fastest way to reach me is email.