AI engineer, Python backend. I build agentic AI systems with Strands Agents, FastAPI, and local or cloud LLMs (Bedrock, Ollama, LM Studio). The agent calls the tools; plain Python handles the scoring and the rules.
馃搷 Chennai 路 馃帗 CSE @ SRM University 路 馃捈 Open to AI Engineer / Python Backend roles 馃敆 Portfolio 路 LinkedIn 路 X
| Project | What it shows |
|---|---|
| mark-map | Multi-agent class-ops desk (AWS Agents for Humans). Turns an exam paper and marksheet into a chapter map and briefs for teachers and families. FastAPI + Strands, Bedrock/Ollama/LM Studio, OCR, CI, and an eval gate that keeps invented marks at 0. Demo video |
| family-loop | Everyday agent desk around Screen Time: triages requests, asks the parent only when it should, and sends a Sunday digest. Strands specialists, policy JSON, persistent memory. Demo video |
| Kasparro-ai-resume-screening | Deterministic resume ranking CLI. The LLM and GitHub API only supply evidence, with a fallback for each resume. pydantic, pytest, PDF reports. |
| openclaw-nexus | Homelab agent system on OpenClaw: an old phone as the camera, a tablet as the kiosk, and a local 20B model, all controlled from WhatsApp. |
| AI-powered-code-analysis-assistant | AST and embedding-based static analysis with refactor suggestions. |
| BankEngine | High-concurrency banking engine in Java, with tests and benchmarks. |
AI / agents: Strands Agents, LiteLLM, OpenAI-compatible APIs, AWS Bedrock, Ollama, LM Studio, Hugging Face, scikit-learn Backend: Python, FastAPI, pydantic, pytest, Docker, GitHub Actions Also: TypeScript / Next.js, Cloudflare Workers, Java
- Guardrails first: missing data stays missing, and no silent yes on sensitive actions.
- Tests and eval checks alongside the code, and demos that run in one command.