Most people learn about AI. I build with it.

I design and ship systems where AI is the architecture, not a bolt-on feature — retrieval pipelines with real memory constraints, multi-agent systems with hard safety boundaries, and satellite-data ML models that run in production instead of sitting in a notebook. My work spans three fronts:

• 🧠 GenAI application engineering — multi-provider LLM abstraction, RAG-style tiered memory, agentic safety design, real-time ASR pipelines
• 🛰️ Applied ML on real-world sensing data — satellite/weather data fusion, physics-informed risk modeling, anomaly detection at production SLAs
• ⚙️ Full-stack + MLOps discipline — Docker, CI/CD, async pipelines, monitoring, and test coverage that treats ML services like real infrastructure, not demos

I care about responsible, guard-railed AI — systems that know what they don't know, escalate to humans when it matters, and never fabricate confidence they haven't earned.

A real-time early-warning system that fuses 6 independent satellite/weather sources (NASA GPM, NASA SMAP, ESA Copernicus DEM, Sentinel-2, NOAA GFS) into a probabilistic Failure Probability Index (FPI), computed at 1 km² resolution and refreshed every 6 hours with 24–48 hour forward forecasting.

• Physics-based risk model derived from NASA LHASA v2, calibrated with a LightGBM classifier and validated against 6 historic landslide events — 6/6 correctly flagged
• FastAPI + async SQLAlchemy + PostgreSQL backend, Celery-scheduled satellite ingestion, Redis caching, Next.js 14 + MapLibre GL live dashboard over WebSockets
• Government-standard CAP v1.2 XML alert feed (NDMA Sachet-compatible) and WhatsApp Business API dispatch with HMAC-verified webhooks — alert-to-delivery in under 30 seconds
• 100+ pytest cases, GitHub Actions CI/CD, Dockerized multi-service deployment behind Nginx, Prometheus metrics
• Sub-120ms p95 API latency; a full-India model run completes in ~4 minutes

Built to close the gap that cost 420 lives in the 2024 Wayanad disaster — a warning existed 16 hours prior but never reached the right channel. SlopeSense wires the model directly into the alert channel.

A dual-signal remote-sensing system that surfaces dated, defensible anomalies — never confirmed violations — along 4 monitored river hotspots (Chambal, Yamuna, Ken, Ganga), built for human-in-the-loop review by journalists, NGOs, and lawyers.

• Sentinel-1 SAR backscatter log-ratio change detection (10·log10(incident/baseline), +3dB threshold) — works through cloud cover and at night
• Sentinel-2 NDWI sandbar-area differencing to estimate physical sand-volume loss, fused into a single anomaly scorer
• Rolling 12-month seasonal baseline to correct for monsoon vs. dry-season bias
• Evidence-card PDF/JSON export pipeline for legal and journalistic use; static JS/HTML dashboard with 30s auto-polling plus a Streamlit interface
• 21 passing unit tests; guardrails baked into the design — never claim "confirmed illegal," never fabricate numbers, always require human review

IEEE CodeBhoomi 2026 — Top 10 National Finalist, Agentic Autonomous Systems track. A palm-sized AI companion built around persistent memory and warmth instead of transactional Q&A.

• Multi-provider LLM abstraction layer (Mistral, OpenAI, Anthropic, fully local Ollama) behind a provider-agnostic interface, with secure OS-keychain credential storage (keytar)
• Tiered, RAG-style memory architecture: Layer 1 always-loaded core identity (~500 tokens) → Layer 2 keyword-retrieved episodic memory (top-3 matches/turn) → Layer 3 a 7-day rolling emotional-state window, assembled into a bounded ~800–1000 token context per request
• Automated episodic summarization: an LLM call at the end of every conversation extracts a structured summary, tags, and mood label back into SQLite — a self-updating knowledge base
• Token-by-token streaming over IPC from Electron's main process to renderer, driving a 9-state sentiment-aware sprite animation system
• Privacy-by-design: better-sqlite3 local-only storage, strict contextIsolation + sandbox: true, no Node integration in the renderer

IEEE CodeBhoomi 2026 — Top 10 National Finalist. A three-tier mental wellness platform (self-guided tools → anonymous peer support → licensed therapy), explicitly engineered against unhealthy parasocial dependency.

• Dual-agent architecture with strict separation of concerns: "Disha," a conversational triage agent that runs a 3–5 minute open-ended Pulse Check and routes users across a 5-level severity scale without diagnosing, and "Kavach," a silent, real-time safety-monitoring agent watching for self-harm signals
• Kavach generates structured post-session learning briefs for student companions and warm-handoff summaries for licensed psychologists
• A concrete example of designing safety-critical, role-separated multi-agent systems rather than a single monolithic chatbot

A Windows system-tray app for global voice dictation, transcribing speech to text anywhere on the OS.

• Groq Whisper API for near-instant ASR (<500ms), with an additional LLM polishing pass (Groq LLaMA) across three configurable output modes: Raw / Semi-formal / Polished
• Custom voice-activity detection (RMS threshold filtering) and a symbol-expansion text-normalization layer
• Python state-machine core orchestrating mic capture, global hotkeys, a PyQt6 dashboard, and pyautogui keystroke injection — packaged into a single self-contained Windows executable via PyInstaller

A fully automated, zero-infrastructure daily newsletter delivered to each subscriber's local 6am.

• RSS ingestion → Mistral AI multi-document curation/summarization across three categories (news, tools/models, funding) → a higher-order "Signal" step that synthesizes cross-category patterns → HTML email delivery
• Orchestrated by an hourly GitHub Actions cron job with timezone-aware logic, Resend for transactional email — fully serverless at $0/month

A fast, aggregated directory of global and India-specific hackathons.

• AI enrichment pipeline (Mistral AI) that extracts structured fields — real deadlines, prize pools, descriptions — from raw hackathon listings: unstructured-to-structured extraction via LLM
• Turborepo monorepo: Next.js 14 frontend with SSG for near-instant loads, Prisma-backed shared DB package, a separate Python enrichment pipeline run via GitHub Actions, deployed on Vercel

GenAI / LLM Engineering Multi-provider LLM abstraction · RAG / retrieval-augmented architecture · agentic & multi-agent safety design · prompt engineering · context-window management · ASR (Whisper via Groq) · LLM-driven summarization & structured extraction

Applied ML & Geospatial Model calibration · retrospective/backtest validation · anomaly detection · Google Earth Engine · Sentinel-1 SAR / Sentinel-2 optical processing · NDWI & SAR backscatter analysis · NASA GPM/SMAP · Copernicus DEM · physics-informed risk modeling (FPI / LHASA v2)

Languages & Frameworks SQL (CTEs, window functions, joins) · Streamlit · Tailwind CSS · shadcn/ui

Infra, Data & MLOps SQLite (better-sqlite3) · Celery · Alembic · Nginx · Vercel · Firebase · Prisma · CI/CD pipelines & cron orchestration

Testing & Analytics Pytest (100+ test suites on SlopeSense, 21 on River Watch) · unit/integration testing discipline · Power BI · Matplotlib · Seaborn · A/B testing · funnel & cohort analysis · revenue forecasting

Creative Tools Premiere Pro · After Effects · Photoshop · Canva · ImageFX · Veo 3

Bluestock Fintech — Software Development Engineer Intern (Feb 2026 – Apr 2026) Full-stack development on a production-ready corporate blog platform — MERN stack, Next.js SSG/ISR, authentication, CMS features, SEO-structured metadata, database design, and CI/CD — in an Agile, remote team environment.

SalesMonk.ai — Operations Management Intern (May 2025 – Nov 2025) Built internal dashboards and reporting pipelines; supported process optimization through data analysis.

Jivika — Marketing Intern (Jun 2025 – Aug 2025) Analyzed marketing campaign performance and content engagement metrics; produced supporting video content.

• 🏆 Top 10 National Finalist, IEEE CodeBhoomi Tech for Good Hackathon 2026 — team "html," project DeTalks
• Responsible AI: Applying AI Principles — Google
• Data Analyst 101 — Microsoft
• Finalist, i-Hack Hackathon (E-Summit '25) — E-Cell, IIT Bombay
• Participant, Google Solution Challenge — Google Developer Groups India
• Participant, EUREKA 2024 — IIT Bombay