BE Artificial Intelligence & Data Science · Pune, India

Selected projects · Tools · Current focus · Contribution snake · Arcade Zone · Mission 2030

I'm a second-year AI & Data Science student building AI applications, backend systems and reproducible data pipelines.

I care about the engineering behind the demo: authorization, data quality, evaluation and failure handling. My long-term direction is AI infrastructure; my current work is about getting the foundations right.

Turn real-world tasks into quests, XP, gold, attributes and streaks, supported by an AI companion called the Oracle.

• Engineering focus: server-authoritative rewards, transactional progression and multi-provider LLM routing.
• Design choice: deterministic code governs critical state; the LLM supports the conversational layer.
• Recognition: Tech Zephyr 4.0 Grand Finale Finalist, IIT Bhubaneswar.

Stack: Next.js · TypeScript · PostgreSQL · Supabase · Drizzle ORM

A three-round data science pipeline for AARUUSH '26: restore corrupted data, evaluate NLP models and track signals in a live-collected corpus.

• Restoration: 12,360 raw rows → 10,221 analysis-ready records, with 27 logged repairs.
• Evaluation: reported macro-F1 of 0.6134 for sentiment and 0.8051 for topic classification.
• Collection: 10,384 records across 21 sweeps and 5 languages.

Engineering focus: traceable repairs, fixed splits and seeds, automated checks, and explicit error analysis. These figures describe the documented project runs; reproduction steps and limitations are in the repository.

Stack: Python · pandas · scikit-learn · TF-IDF · SQLite · Jupyter

A Hinglish-oriented voice assistant exploring basic health information and healthcare facility discovery through natural conversation.

• Pipeline: speech-to-text → language model → text-to-speech, connected through LiveKit.
• Engineering focus: caller memory, facility lookup and specialist handoffs.
• Scope: a health-information prototype, not a diagnostic, prescribing or emergency-response service.

Stack: LiveKit Agents · Gemini · Deepgram · Murf Falcon · SIP · SQLite

Built on the Murf/LiveKit starter, with Anaya-specific voice and health-access workflows.

Hospital discovery and appointment workflows built around a question: where did this information come from, and how fresh is it?

• Engineering focus: per-fact provenance, relationship-scoped access and explicit stale or unavailable states.
• Workflow focus: appointment requests, staff-scoped operations and server-side state transitions.
• Status: demo-backed flows are distinct from verified live hospital integrations.

Stack: Next.js · TypeScript · Supabase · PostgreSQL · RLS · Zod · Vitest

Demo · Upstream project

Public fork of slicqq/FlowCare; the upstream project and this fork's changes are linked separately.

🗂️ More work & prototypes

• TRAXIS — A team project/concept for multi-tenant polar-expedition logistics, assets and inventory. Focus: JWT verification, RBAC and tenant-scoped operations. React · FastAPI · PostgreSQL. Smart India Hackathon 2026 work; no public repository linked here.
• InnerLoop — A separate student-productivity project for planning, focus, habits and analytics. React · Vite · Supabase. Foundational workflows are implemented; the full AI reasoning layer is planned. This is separate from Questbound.
• SHIELD-COM — An embedded speech-processing concept for noisy environments. ESP32-S3 · ESP-SR. Smart India Hackathon 2026 team/internal-round work; intended architecture, not a production-validated system.
• NoLimi — A Python desktop-assistant prototype combining voice interaction, Gemini, browser automation and application launching.

Systems in motion — illustrative architecture

• Strengthening ML foundations and PyTorch through practical model workflows.
• Learning RAG, agentic systems and LLM evaluation, including failure analysis.
• Improving backend reliability through authorization, testing and better data modeling.

• The server owns critical state. Permissions and rewards should not be decided by the browser.
• Claims need evidence. Data and metrics should be traceable, reproducible and accompanied by limitations.
• Failure modes are part of the product. Validate inputs, test boundaries and surface errors clearly.

📊 GitHub activity

Activity graphics are generated by third-party services and may occasionally be unavailable or delayed. They show activity, not a measure of code quality.

Built from my public contribution calendar. Calendar levels are a dated snapshot; the snake's motion is decorative. The export date is shown inside the graphic.

Classic walls kill, No Walls wrap, Obstacles 5 blocks, Portal teleport gates. Easy/Normal/Hard/Insane speeds + power-up P that slows 3 sec.

Play: games/snake.html

Easy 4×4, Normal 4×5, Hard 6×6, Blitz 60s timed, Extreme 90s + 40 moves limit.

Play: games/memory.html

Normal balanced, Hardcore 2× slower 2× gold, Speedrun race to lvl 25, Boss Rush boss every 5 levels, Idle passive focus. Prestige resets for ×1.5 multiplier.

Play: games/quest.html

Easy 3 chars, Normal 4, Hard 5, Extreme 6, Timed 60s, Mastermind no repeats. Hint reveals one char (costs attempt). Streak tracking.

Play: games/hacker.html

Merge tech stacks: PY → TS → PG → RE → FA → SB → DO → LX → K8 → AI → 2048. Undo move.

Play: games/2048.html

Classic Tetris — stack cleanly. Each 10 lines = level up (faster). Next piece preview.

Play: games/tetris.html

Minesweeper — Easy 9×9/10 mines to Extreme 20×20/80 mines. Left reveal, right flag.

Play: games/minesweeper.html

Flap through { } [ ] brackets. Easy wider gaps, Hard narrow, Insane tiny, Reverse flips gravity every 5 points.

Play: games/flappy.html

Squash green bugs 🐛, avoid red features ✨ in Hard/Insane. Normal/Hard/Insane/Zen modes. Accuracy tracking.

Play: games/bug.html

Type code snippets: JS/TS, Python, SQL, Mixed, Hard no-backspace. WPM + accuracy, 60s timer.

Play: games/typing.html

Questbound dungeon — move with arrows/WASD, collect 💰, fight 👾, reach 🚪. Normal/Hard/Rogue permadeath/Explorer 12×12.

Play: games/dungeon.html

Endless runner — jump over bugs, collect XP orbs. Easy/Hard/Insane/Double Jump (2 jumps in air).

Play: games/runner.html

All 12 games are original, client-side only, no trackers, no external APIs. Scores saved locally. Upload the games folder with your README and enable Pages in Settings → Pages → Deploy from main branch.

┌── kunal@mission-2030
│
├── focus       applied AI · backend systems · data integrity
├── loop        build → measure → learn → ship
├── north star  AI infrastructure
└── status      learning, building, improving

Build toward AI infrastructure through strong foundations, useful products and measurable progress.

Open to collaborating on AI applications, backend systems, voice interfaces and data pipelines.

Connect on LinkedIn