Planner & Researcher
A multi-agentic system built using google-adk framework and MCP tools. It searches for new job openings, extracts information like tech-stack, contact info and company details. Agents communicate via state/ tool-contexts and the system implements callbacks, middleware for HITL, for certain actions like save to DB or apply directly. It uses custom tools built on top of MCP to interact with Notion DB.
Google-ADK
State Sharing
A2A, MCP
Multi Agent
Health Dashboard & Trends
Multi-agent system to upload health reports, generate KPI trend graphs, and query for insights. Optimized context window via adk-artifact-manager to reduce token overhead. Evaluated cost reductions using Ollama, DeepSeek, LLaMA. Evaluated performance using prompt versioning and Opik evaluations.
Google-ADK
ADK-Artifacts
Opik Evals
Ollama
Multi Agent
Website RAG Bot
RAG system for portfolio site with heuristic and semantic feedback, golden datasets, and eval-based development. Ingestion and query pipelines, P50/P90 latency tracking, LLM-as-judge feedback metrics, central prompt versioning, and evals gated on every commit via pre-commit hooks and CICD.
RAG
Opik Evals
CICD
LLM-as-Judge
Pinecone
Resiliency Automation Framework
End-to-end game-day automation for Chase Travel — orchestrating load injection, chaos injection, monitoring, and reporting to reclaim 5–6 hours per engineer per day. An AI layer was added to synthesize experiment results into structured summaries for compliance documentation.
AWS
DB-Streams
Lambda
FIS
Gremlin
N8N Email Assistant
Automated email triage workflows using N8N deployed on GCP Cloud Run. Daily spam cleanup, inbox prioritization, and unread mail summary sent to Slack via cron jobs. Deployed with N8N Community Docker using Secret Manager for API keys.
N8N
GCP Cloud Run
Docker
Slack
SLMs for Edge Devices (In Progress)
Building a voice-enabled autonomous phone assistant using PyTorch-trained SLMs for Android. Will train on recorded calls to classify spam/ham and answer autonomously. Exploring Android NDK + llama.cpp integration and STS models to reduce STT/TTS latency.
PyTorch
SLM
Android NDK
llama.cpp
In Progress
Comfy UI Setup On AWS
I setup a comfy UI rig on an EC2 instance to play around with image generation using open source Flux, SDXL models and LoRAs. While setting up manually was excruciating, it was a good learning curve.
Stable Diffusion
ComfyUI
AWS
WhatsApp Companion Ava
Production-ready multimodal app (TTS, STS, TTI) with LangGraph backend, WebRTC and FastAPI endpoints. Real-time WebRTC audio streaming with open-source STT/TTS/TTI models. Context-aware memory using summarization and user-preference extraction stored in Qdrant to reduce drift.
LangGraph
WebRTC
FastAPI
Qdrant
TTS/STS/TTI
Media Server On Raspberry Pi
Raspberry Pi w USB setup to host NFS server on home LAN. Allows downloaded media (TV shows, Movies) to be copied to attached USB storage via phone, and can eventually be played on smart TV.
Raspberyy Pi
NFS Server
SSH
Termux