Résumé

Résumé

AI SaaS · May 2026 – Present

Solo founder & full-stack engineer — web, backend, ML pipeline, 3 native clients, infra

Bot-free meeting-intelligence SaaS — multi-engine speech-to-text, AI summaries and RAG semantic search across Chrome, macOS, iOS and web on one distributed backend.

• Architected a production meeting-intelligence SaaS spanning four client surfaces — a Manifest V3 Chrome extension, a Tauri v2 + Rust macOS app, an Expo / React Native iOS + Android app and a Next.js 16 web vault — all sharing one distributed TypeScript backend that turns raw meeting audio into editorial summaries, action items, follow-up drafts and cross-meeting reports.
• Built a multi-engine speech-to-text pipeline with automatic failover across Google Vertex AI Gemini 2.5 Pro, Groq Whisper-large-v3 and Google Cloud STT v2 Chirp 3 — FFmpeg transcoding to 16 kHz FLAC, dual-channel speaker diarization, Hindi/Hinglish code-switch handling and a cost circuit-breaker so the pipeline never hard-fails.
• Implemented RAG Q&A and semantic vault search on PostgreSQL/pgvector — Gemini embeddings (768-dim) ranked by cosine distance over an HNSW index with transparent keyword-search fallback, and summarization on Gemini with a LLaMA 3.3-70B (Groq) fallback.
• Offloaded long-running synthesis to a Dockerized Google Cloud Run worker dispatched via Google Cloud Tasks with OIDC service-to-service auth — a dark-launchable switch escaping Vercel's 300s function cap, with idempotent job dedup, status-mapped retries and a daily Cron re-enqueue for self-healing.

iOS App · June 2026

Solo iOS engineer — design, architecture, App Store release

Native iOS anime-tracking app on the App Store — Expo SDK 56, React Native New Architecture, iOS 26 Liquid Glass, a pure REST client of a Next.js backend.

• Built a native iOS app published to the Apple App Store as a pure REST client of a Next.js backend — a typed apiFetch wrapper over 20+ /api/* endpoints authenticating via the backend's httpOnly session cookie replayed by the iOS cookie jar, with no Authorization headers or manual cookie handling.
• Architected server state with TanStack Query v5 — infinite-scroll paginated library lists and fully optimistic add/edit/delete/complete/reorder mutations that transform every mirrored cache in onMutate to eliminate stale-flash, with pure unit-tested cache transforms.
• Engineered a fully native iOS 26 Liquid Glass UI (expo-glass-effect) that degrades to BlurView below iOS 26 — across the tab bar, headers, sheets and toasts — with Reanimated 4 springs, haptics, blurhash posters, skeletons and pull-to-refresh.
• Shipped a full social graph (search, profiles, follow/followers, block/report), timezone-aware airing countdowns (JST→IST), search-to-add via the Jikan/MyAnimeList API, one-tap .xlsx export to the iOS share sheet, and a RevenueCat lifetime in-app purchase keyed to the backend user.

AI SaaS · October 2025 – June 2026

Solo full-stack engineer — web backend, AI engine, iOS app

NLP-driven class-attendance app on the App Store, powered by a Groq LLM engine that parses plain-English commands across hundreds of student records.

• Built the core AI attendance engine as a Next.js Route Handler wrapping the Groq SDK — a system prompt feeds the ordered roster plus current attendance state to an LLM that returns a strict JSON object of enrolment IDs to mark present/absent (temperature 0, JSON mode), so a class rep updates attendance with one plain-English sentence.
• Engineered a resilient 6-model Groq fallback chain (GPT-OSS-120B, LLaMA 3.3-70B, Qwen3-32B, LLaMA 4 Scout/Maverick, LLaMA 3.1) that retries only on HTTP 429 rate limits and fast-aborts on auth errors — keeping the AI feature available across free-tier limits (~99% uptime).
• Prompt-engineered advanced parsing: spatial/positional range commands ('everyone up to Ranvir is present') resolved against the array's physical order, and mathematical-logic commands ('roll number less than 3050') evaluated per student from numeric tokens.
• Shipped two clients on one shared REST backend — a Next.js (shadcn/Radix + Tailwind v4) web app and an iOS Expo / React Native app with Apple Liquid Glass, JWT (jose) + Google/Apple JWKS auth, and EAS Build/Submit to the Apple App Store.

Civic Tech · Geospatial Analytics · June 2026

Solo full-stack engineer — frontend, geospatial mapping, Python data pipeline & ML

Civic-tech parking-congestion intelligence turning ~298,000 Bengaluru Traffic Police tickets into a live hotspot map and a deployable, phone-ready patrol schedule.

• Built a map-first congestion-intelligence dashboard (Next.js 16, React 19) where a deck.gl + MapLibre GL stack renders a heatmap density surface plus hotspot, congestion and blind-spot layers over a MapmyIndia vector basemap, with automatic CARTO raster fallback so the map always renders.
• Engineered a Python pipeline that ingests ~298,000 Bengaluru Traffic Police parking-violation records and emits precomputed JSON aggregates, scoring each corner by a 1–5 carriageway-obstruction severity binned onto a ~220 m grid so a blocked arterial outranks a quiet side street.
• Trained a scikit-learn spatio-temporal forecaster — an ensemble of RandomForestRegressors (200 trees) blending count- and log-scale predictions with cyclical hour/day features — to predict violation intensity per grid cell per hour-of-week (R² ≈ 0.65 on a held-out split).
• Generated deployable patrol beat sheets per station/shift via a greedy nearest-neighbour TSP over haversine distances, with a no-canvas inline-SVG QR handoff to a constable phone mode (live Geolocation distance, offline localStorage persistence) — shipped as an installable PWA.

On-Device ML · Edge AI · May – June 2026

Solo ML + mobile engineer — Hackathon 7.0

Offline, on-device face recognition + anti-spoofing liveness for field personnel — three ONNX deep-learning models, zero network at inference.

• Engineered a fully offline, on-device biometric pipeline running three ONNX deep-learning models with zero network calls at inference — YuNet face detection, MobileFaceNet ArcFace 512-D embeddings (cosine matching) and MiniFASNet-V2 passive anti-spoofing — on one Expo / React Native codebase for both iOS and Android.
• Re-implemented the entire computer-vision pipeline in TypeScript: ported the YuNet output decoder bit-exact against OpenCV (0.000 px error) and hand-wrote a 4-DOF least-squares similarity solver for 5-point ArcFace face alignment to 112×112.
• Built a CSPRNG-randomized active liveness challenge–response system (head-left / head-right / smile gestures derived from facial landmarks) that defeats both printed-photo and video-replay spoofing because the attacker can't predict which gesture will be commanded.
• Authored a Python quantization harness (ONNX Runtime dynamic int8, fp16 fallback) that compressed the recognition model 72% (4.80 MB → 1.36 MB) while retaining a 0.89 cosine separation margin — total shipped model footprint 3.33 MB.

Web App · January 2025 – Present

Solo developer / frontend engineer

Official React + Vite site for Amity University Jharkhand's NSS unit — events, special camps, gallery, team and volunteer enrollment.

• Built the official website for Amity University Jharkhand's NSS (National Service Scheme) unit as a React 18 + Vite SPA styled with Tailwind CSS — 18 routes spanning About NSS, special camps, media gallery, program officers, events, volunteers and a joining form.
• Engineered a responsive mega-menu from scratch with multi-level dropdowns, scroll-triggered sticky state, window-resize handling and click-outside dismissal, plus an auto-advancing hero carousel with opacity cross-fade — no carousel dependency.
• Implemented per-page code-splitting via React.lazy + Suspense and config-driven routing (react-router-dom 7), deployed on Vercel with SPA rewrites under a custom subdomain.