Profile

Kushal Panchali

AI Engineer

Visakhapatnam, India
GitHubLinkedInEmail

Experience

AI Engineer Intern

Peppo AI

Hyderabad, IndiaSep 2025 – Mar 2026
  • Built an offline AI voice assistant on Raspberry Pi, connecting microphone input, speech recognition, local LLM response generation, and text-to-speech without cloud APIs.
  • Improved response time from 20–25s to 4.48s by profiling the pipeline, tuning execution flow, and improving audio buffering on limited hardware.
  • Added live local-model switching without restarting the system, making experiments and model comparisons faster.

Contributor

Terax AI

Open SourceApr 2026 – Present
  • Contributed to an open-source desktop ADE combining terminal, editor, file explorer, web preview, and assistant-driven development workflows.
  • Worked on AI tooling interaction flows, including terminal and editor context for local assistant-assisted development.

Projects

miea-mem - Graph Memory for AI Agents

Graph memory for AI agents as readable JSON triples. No extraction pipeline, signpost steering at read time.

PythonMCPClickPyPI

AI-Native HRMS Portal

Full-stack HRMS portal with role-based access, AI resume screening, interview summarization and automated HR requests.

Next.jsSupabaseAWS S3NVIDIA API

Offline Voice Assistant on Raspberry Pi

Microphone input, speech recognition, a local LLM and text-to-speech combined into one working system on limited hardware.

VoskOllamaWhisperPiperPython

Kliae - Graph based learning environment

Open source graph-based learning platform that organizes knowledge as interconnected concepts instead of linear notes.

TypescriptZustandReact FlowTauri

Database Security Layer for AI Agents

Python asyncio microservice that parses the PostgreSQL wire protocol at the TCP layer and blocks destructive queries before execution.

PythonAsyncioFastAPIDockerAWS ECR

Adaptive Multi-Agent Workflow Monitor

Built an adaptive policy engine that dynamically steers coder and reviewer agents by injecting runtime guidance based on detected workflow signals, attempting recovery before terminating inefficient execution.

AWS CloudWatchLangGraphNext js