I build things with code. RNNs that remember. CNNs that see. ANNs that approximate. Machine learning, AI automations, productivity tools, anything that saves someone an hour, a day, a week of their life.
Production incidents used to mean someone opening ten dashboards at 2 a.m., trying to piece together what broke and why. I built a multi-agent system on LangGraph that does that investigation itself, pulls logs, correlates signals, reasons through the incident, and hands back a root cause instead of a pile of raw data.
Mean-time-to-resolution dropped by ~60% across critical workflows. The best compliment it got was an engineer asking why the alert already had an answer attached.
Started small: a sample dataset, a hunch that anomaly detection could be automated instead of eyeballed. Ended up extending the same pipeline to ingest live production network telemetry, using Qdrant and agentic RAG to flag anomalies as they happen instead of after someone complains.
The gap between "works on sample data" and "works on the real thing" is where most projects die. This one didn't.
This one had actual business risk attached: get it wrong and a customer gets a bad quote. Built with LangGraph routing and tool-calling, it answers customer queries, generates real service quotes against live pricing APIs, and tracks orders all without a human touching it.
Response time went from 8 hours to under 5 minutes. That's not a rounding error, that's the difference between a customer waiting and a customer buying. Retention and acquisition both moved because of it.
Training a model is the easy 20%. I wanted the other 80% proper experiment tracking with DVC and MLflow, a FastAPI endpoint on AWS, and a feature that compares a listing against 50 nearby ones so a host knows why the price is what it is, not just what it is.
An AI tutor that answers questions on NCERT Class 12 English, poem by poem. Built vector search with Qdrant and hierarchical indexing with LlamaIndex so it retrieves at the level of a single poem, not a whole chapter. Groq's LLaMA 3.3-70B handles the answers, wrapped in a Gradio UI.
Small project, but it taught me more about retrieval granularity than any tutorial did.
Give it one topic. A Researcher agent digs it up, an Editor agent shapes it, a Designer agent lays it out built on CrewAI, running on LangChain and LLaMA 3.3-70B. Out comes a finished HTML newsletter with live status tracking.
Watching three agents hand work off to each other without stepping on one another is still, months later, one of my favorite demos to run.
Fine-tuned EfficientNetB2 on Food-101 — 101 categories, 7.8M parameters — landed at 89% accuracy. Deployed on Gradio with top-3 predictions and confidence scores, because "probably a burrito" is more honest than a single overconfident guess.
Agentic AI & Orchestration
LangGraph LangChain LlamaIndex CrewAI DeepAgents MCP Agentic RAG Knowledge Graphs
ML & Deep Learning
PyTorch TensorFlow Hugging Face Transformers LoRA/QLoRA CNNs NLP Computer Vision XGBoost
Voice AI
LiveKit Agents WebRTC STT/TTS Deepgram ElevenLabs VAD SIP/Telephony
MLOps / LLMOps
MLflow DVC LangSmith Docker Kubernetes CI/CD Grafana Prometheus FastAPI
Vector Stores & Data
Qdrant Milvus Chroma FAISS pgvector Redis PostgreSQL
Cloud & Languages
AWS Azure OpenAI Azure ML Python SQL
IBM RAG and Agentic AI · IBM Generative AI Engineering · Microsoft AI & ML Engineering · DeepLearning.AI — NLP & Deep Learning · Model Context Protocol Mastery · IBM AI Engineering