Software engineering student working across AI, machine learning, and full-stack systems — with a bias for civic-tech that stays transparent, deterministic, and grounded rather than opaquely AI-generated. Five production-grade builds, solo, end to end.
I'm an Information Science and Engineering student at Ramaiah Institute of Technology, building full-stack and AI-driven applications across the entire SDLC — design through deployment. My stack runs through C and C++, Python, and JavaScript/TypeScript, REST APIs and WebSocket services, and containerized infrastructure with Docker, Docker Compose, and GitHub Actions CI/CD.
I'm particularly drawn to India-focused, socially impactful civic-tech: healthcare-cost transparency, government paperwork, personal-finance intelligence, air-quality health risk, and career readiness. The common thread across my projects is keeping the logic auditable — deterministic scoring, transparent rules, and LLMs used only where they genuinely help, never as a black box standing in for a decision.
Outside of engineering, I read and write — poetry, prose, and the occasional technical blog post.
CGPA: 9.51 / 10
Percentage: 97.67%
Percentage: 95.40%
Currently contributing to ongoing project work as part of this internship.
Contributed across the full SDLC of an AI Disaster Prediction and Response System: a FastAPI/Python REST backend with WebSockets, SQLAlchemy, and JWT auth, paired with a React.js/Vite frontend using Tailwind CSS and React Leaflet.
Wrote articles on social issues and possible solution strategies, and supported organizational fundraising initiatives across social platforms.
Compare a resume against real market demand for a target role and return a defensible job-readiness score and roadmap.
Deterministic skill matching and scoring over a curated 18-role job-market snapshot (live Adzuna/JSearch optional), with a ranked learning roadmap carrying hour estimates and .ics calendar export; Groq writes narrative text only. Full auth stack — rotating refresh cookies, email verification, password reset, rate limiting, account lockout, data export.
Scored skill-gap reports, role comparison, PDF export, and shareable read-only links — deterministic templates keep it fully functional with no LLM key set.
Tie a person's own lagged pollutant exposure to their own symptom pattern — honestly, without pretending the data is cleaner than it is.
An exposure-window pipeline that aligns three irregular time series and stores an explicit data_coverage_pct per window; lagged features at t-0 / 6h / 24h / 72h feeding a per-user RandomForest, with an anonymized population fallback below a minimum-data threshold; a pgvector RAG layer that only rephrases cited WHO / CPCB passages and refuses when retrieval is weak.
Strict time-based train/test evaluation, feature importances exposed via GET /risk/explain, and query-layer per-user isolation verified by tests.
Show what a treatment should cost in India — and whether a hospital bill is fair — before you pay, with no figure ever generated by an LLM.
A five-tier rule-based cost engine (city + hospital type → city → state pool → cited national reference → sample average), each response labelled with the tier that produced it; deterministic scheme-eligibility rules across PM-JAY, CGHS and 7 state schemes; a government-first hospital finder; and a Groq vision layer doing OCR-only on bill photos while the fairness verdict stays rule-based. 8-language i18n with compile-time-checked keys.
Tiered estimates with confidence scores, episode estimation, crowd-sourced cost contributions with admin review, and clean degradation with no Groq key; CI runs ruff, Alembic checks and pytest against Postgres.
Take a bank statement in any shape — CSV, PDF, scanned photo, any column order or date format — and return categorized transactions, spending views, anomaly flags, and natural-language Q&A.
Structural-fingerprint bank profiles that learn a layout once and skip cold detection thereafter; three-tier categorization (exact rule → RapidFuzz → batched LLM fallback) with a promotion job that turns repeated LLM decisions into permanent Tier-1 rules; a per-user IsolationForest with plain-text explanations; and "ask your finances" as text-to-SQL over reviewed parameterized templates — the LLM only picks a template and fills typed params, never emits SQL.
100% deterministic categorization and 1.0 anomaly recall on the seeded set, refresh-token rotation with reuse detection, a per-user audit log, and ~146 tests.
Turn a government notice, scheme or form — PDF or a photo of the page — into a plain-language summary, eligibility checklist, key dates and rebuilt tables, in 7 Indian languages.
No Node backend and no Redis — the job queue is a Postgres table with FOR UPDATE SKIP LOCKED, the worker is a Python process running an extract → simplify → embed → tables pipeline, and live progress streams over WebSockets via Postgres NOTIFY. Multimodal extraction goes through Groq vision with Tesseract OCR as an offline fallback; a document-scoped RAG chat answers only from the uploaded file.
End-to-end pipeline with live per-stage progress, JWT auth with rotating refresh tokens, and an LSA / pgvector retrieval layer that keeps answers grounded in the source document.