2nd-year B.Tech AI & ML student at Atlas SkillTech, Mumbai. I build things that ship, like RecSys on 33M records, audio intelligence platforms, RAG pipelines, computer vision apps.

Currently: shipping production tools — hiring pipeline, order tracking, admin dashboards — for a manufacturing company.

• rtk-ai/rtk #2986 — Replaced live HTTP calls in the benchmark suite with local fixtures for deterministic, flake-free runs.
• LobsterTrap/lola #220 — Added support for lola.yml as an alternate module-config filename, preserving lola.yaml precedence.
• dariushoule/x64dbg-automate-pyclient #21 — Fixed a connection leak: the client now closes its ZMQ connection when a plugin-version handshake fails.
• livekit/agents #6398 — Fixed Gemma reasoning-marker tokens leaking into TTS/transcripts/chat history; added a stateful streaming filter that handles markers split across chunks.

Context-aware recommendation system on MovieLens (33M ratings · 86k movies · 330k users). Three-stage pipeline: TF-IDF content-based filtering, SVD collaborative filtering, configurable hybrid (70/30). Evaluated with RMSE, MAE, Precision@K, Recall@K, NDCG@K on 80/20 split. FastAPI backend with 8 REST endpoints, TMDB poster caching, 0.08–0.25s response time.

Upload a call recording, get back a full transcript, speaker breakdown, sentiment timeline, and AI coaching output. Whisper transcription at 95%+ accuracy, KeyBERT keyword extraction, VADER + HuggingFace emotion timeline, Gemini coaching. Chart.js dashboard with real-time sentiment charts, one-click PDF export via FPDF2. pytest + mypy + flake8 CI enforced. Under 2s API response on 100MB audio.

Multi-document RAG pipeline: HuggingFace sentence-transformers + FAISS + LangChain + Gemini. Supports PDF, TXT, and DOCX. Source-chunk citations, conversation memory, Dockerised deployment.

YOLOv8n inference on uploaded images and live webcam feed via Streamlit. Confidence threshold slider, class filter, batch processing, exportable annotated results. Multi-stage Docker build with 60% image size reduction.

Dual TTS engine (Coqui + ElevenLabs) with CUDA-accelerated neural voice cloning. Cold start 2–3s · warmed up ~0.5s · cached phrases under 0.1s. Smart async queue, usage analytics, CLI benchmark measuring P50/P95/P99 latency.

stack = {
    "languages"  : ["Python", "SQL", "Java", "Bash"],
    "ml_dl"      : ["scikit-learn", "PyTorch", "XGBoost", "LightGBM", "YOLOv8", "HuggingFace"],
    "genai"      : ["LangChain", "FAISS", "Whisper", "Gemini API", "Ollama"],
    "backend"    : ["FastAPI", "Flask"],
    "devops"     : ["Docker", "Docker Compose", "Linux (Arch)", "Nginx", "Grafana", "Git"],
    "data"       : ["Pandas", "NumPy", "Matplotlib", "Seaborn", "Plotly"],
    "cloud"      : ["Azure", "GCP", "Streamlit Cloud"],
}

Self-hosted Linux server on Proxmox VE. All services containerised with Docker Compose behind an Nginx reverse proxy.

SSH hardened · 99% uptime across 4+ concurrent services.

• AMD x Google Slingshot Hackathon — #10 / 300
• Mastercard Cybersecurity Simulation · Forage
• 5+ AI/ML Hackathons · Devpost

LinkedIn · Email · GitHub