I build AI agents, machine learning pipelines and backend services. For LLM apps that means the retrieval, tool-routing and safety layers around the model, and I write down what each system cannot do as clearly as what it can.
| 🎓 Studying | B.Sc. Software Engineering, University of Europe for Applied Sciences |
| 🧠 Fellowship | IIT Roorkee Agentic AI Systems & Design cohort 🥈 Excellence in Evaluation (Silver) · 🥇 Total Engagement (Gold) |
| 💼 Open to | Internships in agentic AI, machine learning and backend engineering |
| 🔭 Now | Applying agentic AI, ML and backend skills to real business use cases |
| 🌍 Based in | Berlin |
| Answers come from retrieved documents or real database rows, not from guesses. | Masking, injection checks and schema-validated output sit around the model. | Each system ships with a test set and scores, so its claims can be checked. |
An e-commerce support agent that answers from real policy documents and real order records.
flowchart LR
U([👤 Customer message]) --> C{{🧭 Classify intent<br/>+ mask PII<br/>+ check injection}}
C -->|policy question| R[📚 RAG<br/>ChromaDB]
C -->|order status| S[🗃️ SQL lookup<br/>SQLite]
C -->|feedback| F[💬 Feedback]
C -->|needs a person| H[🙋 Hand off to human]
R --> G[✍️ Schema-validated reply]
S --> G
F --> G
H --> G
G --> O([✅ Response])
| 🔀 Graph | 📄 Knowledge base | 🗄️ Data | 🧪 Evaluation |
|---|---|---|---|
| 6 nodes 5 routing branches |
38 policy PDFs | 50 synthetic orders | 15 test queries |
- 🔒 Phone and card-number masking, prompt-injection detection, JSON-Schema-checked answers
- ♻️ SQLite checkpoints let an interrupted run resume
- 🔌 FastAPI backend and an MCP server
- 🔑 Runs with no API key (
MOCK_LLM=true). All data is synthetic.
|
A support bot where one orchestrator agent delegates to a RAG agent, a text-to-SQL agent, and feedback and escalation tools. Prompts and model settings are versioned files. |
🏛️ Sindhu-RAGA configurable RAG pipeline over a 200+ page corpus on the Indus Valley Civilization, with config-tracked experiments, versioned prompts and a Streamlit test UI. |
|
Linear vs ensemble models for predicting California property values. |
Threshold tuning for malignant tumor detection, evaluated with a confusion matrix and ROC-AUC. |
|
Pulls stock data from the Alpha Vantage API, cleans it with NumPy and Pandas, and explores it with Matplotlib and Plotly charts. |
|
🤖 Agentic AI & GenAI
🐍 Core languages
📊 Data & ML
🚀 Backend & deployment
🧰 Tools