A proactive AI companion that calls elderly individuals daily — reasoning, remembering, and caring, all without human intervention.

Built an end-to-end agentic voice pipeline using Twilio, Faster-Whisper (STT), Kokoro (TTS), and a local Llama3 model via Ollama — the agent autonomously initiates calls, holds natural multi-turn conversations, and tracks emotional wellbeing over time. Episodic memory (pgvector + sentence-transformers) lets Aria remember past calls and personalize each conversation.

Key Highlights:

• Fully autonomous call orchestration: Twilio → STT → LLM → TTS pipeline with no human in the loop
• Episodic memory via LLM-based fact extraction + semantic retrieval (pgvector cosine similarity)
• Real-time mood scoring from acoustic features (energy, pitch, pause ratio via Librosa) with email escalation on detected distress

A multi-agent system that negotiates meeting times on behalf of participants — reasoning about seniority, constraints, and preferences the way a skilled executive assistant would.

Built a 4-agent LangGraph pipeline where advocate agents propose slots per participant and an arbitrator runs deterministic interval-intersection on 20+ real Google Calendar slots before LLM selection, preventing hallucination. The negotiation loop runs up to 3 rounds before escalating. Backed by a FastAPI backend and React UI.

Key Highlights:

• Deterministic overlap detection before any LLM call — LLM only selects between verified slots
• Google Calendar freebusy API with per-participant OAuth and automatic event creation on consensus
• ~10s end-to-end; 45 tests across overlap logic, graph runs, and state schema; GitHub Actions CI

An agent that explores and understands codebases on its own — no hand-holding required.

Built an autonomous system using local LLMs (via Ollama) that traverses a codebase, infers architecture, and generates human-readable summaries. Dependency graphs are visualized interactively, letting developers onboard to unfamiliar repos in minutes instead of hours.

Key Highlights:

• Autonomous codebase exploration using local LLMs — no cloud dependency
• Architecture summary generation with interactive dependency visualization
• Combines AI reasoning + backend orchestration + graph visualization

A multi-step LLM pipeline that doesn't just find bugs — it reasons through and fixes them.

Designed an agentic system integrating static analysis tools, LLM-based reasoning, and automated validation to detect and patch C/C++ vulnerabilities end-to-end. The pipeline chains multiple reasoning steps: scan → understand → generate patch → validate.

Key Highlights:

• Multi-agent pipeline: static analysis → LLM reasoning → patch generation → validation
• Reduced high-severity vulnerability count by up to 70%
• Demonstrates AI in systems-level security engineering

Full stack observability system built with production infrastructure in mind.

Designed and deployed a log analysis platform using Flask, Docker, Prometheus, and Grafana with full CI/CD integration. Real-time metrics dashboards and alerting pipelines make it a practical DevOps companion.

Key Highlights:

• End-to-end DevOps pipeline: Flask backend + Prometheus metrics + Grafana dashboards
• Containerized with Docker; CI/CD automated deployments
• Production-grade observability from scratch

A cryptographic vault with multi-party unlock — security engineering meets full stack.

Built a system that uses cryptographic primitives to lock/unlock sensitive data requiring multiple authorized parties. Includes a React frontend, REST API layer, and cloud-backed storage — designed with security and usability in balance.

Key Highlights:

• Cryptographic multi-party unlock mechanism
• Full stack: React + REST APIs + cloud storage
• Thoughtful system design balancing security and UX