Technical Skills
Programming & Backend: Python, SQL, SQLAlchemy, FastAPI, Flask, REST APIs, WebSockets, AsyncIO, Celery, Kafka
Generative AI & Agentic Systems: LLMs, RAG, LangChain, LangGraph, LlamaIndex, CrewAI, MCP, Tool Calling, Multi-Agent Systems, Prompt Engineering, SFT, LoRA, QLoRA, Embeddings, LLM Evaluation, OpenAI, Gemini, Claude
AI Systems Engineering: Retrieval Optimization, Vector Search, Latency Optimization, Streaming, Caching, Observability, Vector Databases (Pinecone, ChromaDB, Qdrant, FAISS), Hugging Face Transformers
Machine Learning & Deep Learning: Scikit-learn, PyTorch, TensorFlow, Transformers, CNNs, Transfer Learning, XGBoost, LightGBM, Feature Engineering, Model Training & Evaluation, Hyperparameter Tuning, Computer Vision, OpenCV, SHAP, Time-Series Forecasting
Voice AI: Speech-to-Text (STT), Text-to-Speech (TTS), Streaming Audio, Real-Time Voice Agents, Conversational AI, Voice Activity Detection (VAD), Speaker Diarization, WebSockets, WebRTC, OpenAI Realtime API, Whisper, Deepgram, ElevenLabs, Twilio
Databases & Infrastructure: PostgreSQL, MySQL, MongoDB, Redis, Docker, CI/CD, AWS (EC2, S3), Linux, Git
Selected Projects
AI & Backend EngineerAgentic Trip & Ride Management Assistant
LangGraph, Gemini, FastAPI, MCP, WebSockets, Kafka, Redis, MongoDB, MySQL, Docker
- Overview: A conversational assistant for ride discovery, booking, cancellation and trip-history queries, over text and real-time voice, where transactional data must stay correct.
- Design: Stateful LangGraph workflows validate required inputs and route intents. The LLM calls backend-defined MCP tools and never generates or executes SQL. State is split by purpose: Redis for active context, MongoDB for conversation history, MySQL for ride transactions.
- Integrations: Google Places API for location validation, streaming STT/TTS over WebSockets for voice, and Kafka for asynchronous request ingestion.
- How it was solved: Profiled LLM, tool-call, database and context-loading latency separately and optimized each, reducing average response time from ~6.5s to ~1.8s. Tool-only database access removed the risk of model-written queries.
AI & Backend EngineerHealthcare AI: Voice-Enabled Appointment Booking
LangGraph, LangChain, FastAPI, MCP, Twilio Voice, OpenAI, Deepgram, ElevenLabs, PostgreSQL, Redis, Docker
- Overview: A real-time voice AI agent that handles patient appointment booking, rescheduling and cancellation over phone calls.
- Pipeline: Inbound call (Twilio) → Streaming STT/TTS → LangGraph agent router → Conversation state & memory → MCP tools (EHR, calendar, notifications) → Confirmation → SMS/email.
- Design: LangGraph workflows for intent detection, patient verification, availability checks and multi-step booking, with persistent per-call state. Dockerized FastAPI services on PostgreSQL and Redis hold conversation state and appointments and run asynchronous workflows.
- How it was solved: Gave the agent controlled access to EHR, calendar and notification services through MCP instead of direct system access, and kept per-call state so multi-step bookings stay consistent through the call.
Data ScientistSales Decline Root-Cause Analysis & Forecasting
Python, Pandas, statsmodels, LightGBM, SHAP, Prophet, Streamlit
- Overview: An end-to-end analytics system to find where sales declined, quantify the business drivers and forecast future sales across products, regions and channels.
- Design: Time-series analysis, statistical testing and feature engineering over price, promotion, product, region, channel and temporal variables. LightGBM with SHAP for driver analysis; Prophet for forecasting.
- How it was solved: SHAP attributions quantified which drivers explained the decline, and forecasts reached ~9% MAPE under time-based validation. Results were delivered through an interactive Streamlit dashboard.
Machine Learning EngineerVisual Quality Inspection with Deep Learning
Python, OpenCV, PyTorch, Albumentations, Scikit-learn, FastAPI, Docker
- Overview: An automated visual inspection system that detects product defects from images.
- Design: OpenCV preprocessing and transfer-learning CNNs, with augmentation for lighting, orientation and image-quality variation.
- How it was solved: Reached 0.93 precision and 0.91 recall, and tuned the decision threshold on the relative cost of missed defects versus false alarms. Deployed through a Dockerized FastAPI service with defect-localization overlays.