Sudhamsh Kalakonda | AI/ML Engineer

01

LLM Guardrails Gateway

2026

Middleware for LLM input/output safety, built from scratch as a decorator API that drops into any pipeline regardless of provider.

Proven across both Groq and OpenAI providers. YAML-driven policy engine, full audit logging, and a self-eval harness with documented, honest known limitations rather than inflated claims.

FastAPI Groq OpenAI Policy Engine Audit Logging

LLM Guardrails Gateway architecture diagram

Request flow through the gateway, with blocked requests short circuited before reaching the provider

02

Enclave.AI

2025

A privacy-first, multi-agent enterprise AI platform built for Indian BFSI companies navigating RBI and DPDP compliance.

Built from scratch on Ollama and Qdrant, with BAAI/bge-base-en-v1.5 embeddings. LangGraph orchestrates a multi-agent system across retrieval, reasoning, and compliance agents. FastAPI backend, Next.js frontend.

Ollama Qdrant LangGraph FastAPI Next.js BFSI / DPDP

GitHub ↗

LangGraph orchestrating retrieval, reasoning, and compliance agents to answer a query end to end

GitHub ↗

03

Arigatho.voice

2025 to Present

A production-grade voice AI agent that takes real phone calls for a CA firm, listening and responding in real time across Hinglish, Telugu, and English.

Sarvam Saarika for speech-to-text with automatic language detection, Groq's Llama 3.3 70B for responses, Sarvam Bulbul for text-to-speech. Full 8-phase system design documented end to end. Currently blocked on Exotel's bidirectional audio playback pending KYC/TRAI verification on the trial account.

Sarvam Saarika Sarvam Bulbul Groq Llama 3.3 70B Exotel Telephony

Real-time voice pipeline handling speech to text, LLM response generation, and text to speech across Hinglish, Telugu, and English

05

LLM Eval Pipeline

2025

An automated evaluation pipeline built to understand evals deeply enough to defend design decisions in interviews, not just run a benchmark.

Golden dataset design, heuristic evaluators alongside LLM-as-judge scoring, an async FastAPI pipeline, GitHub Actions CI/CD, and a Streamlit dashboard for reviewing results.

FastAPI LLM-as-Judge GitHub Actions Streamlit

Golden dataset flowing through heuristic and LLM-as-judge evaluators into a live results dashboard, triggered automatically on every run

06

ArigatoAI

2025 to Present

A production RAG chatbot handling tax and compliance queries for a real CA firm. Retrieves from 2,400+ indexed vectors and responds in under 2 seconds, running live for actual clients daily.

Pinecone for vector storage, OpenAI embeddings, Groq's Llama 3.3 70B for generation. FastAPI backend, Next.js frontend, with an embeddable website widget and an admin dashboard, deployed on Render and Vercel.

FastAPI Next.js Pinecone OpenAI Embeddings Groq Llama 3.3 70B Render Vercel

View Live ↗

User query retrieving from indexed vectors and returning a grounded response in under 2 seconds, with admin visibility into every request.

View Live ↗

07

VisAssistance

Apr 2026 to Present

An AI-powered visa assistant for international students navigating F1, OPT, CPT, and H1B questions. Features real-time model switching between Llama 3.3 70B, GPT-4o mini, and Gemini 1.5 Flash, so you can compare answers across providers instantly.

Built a unified LLM client in FastAPI that abstracts Groq, OpenAI, and Google APIs behind a single interface. Persistent chat history, markdown rendering, and a clean sidebar, all deployed on Render and Vercel.

FastAPI Next.js Groq OpenAI Gemini Llama 3.3 70B Render Vercel

View Live ↗

Landing: model switcher and suggested prompts

Active chat: markdown rendering, chat history

View Live ↗

08

風 Kaze

Apr 2026

An open-source AI companion for any website. Drop a single HTML script tag and an animated character (Samurai, Fighter, or Shinobi) appears, walking across the screen and ready to chat.

Built with pure vanilla JavaScript and zero dependencies. It supports multiple AI providers including Groq (Llama 3.3 70B), OpenAI, Anthropic, and Gemini. Fully configurable via data attributes with a strict focus on local privacy (API keys stay in the browser).

Vanilla JS Groq Llama 3.3 70B Open Source

View Live ↗ GitHub ↗

Embed instructions, zero dependency setup

A lightweight companion powered by heavy-hitting AI

View Live ↗ GitHub ↗

Selected Work

LLM Guardrails Gateway

Enclave.AI

Arigatho.voice

LLM Eval Pipeline

ArigatoAI

VisAssistance

風 Kaze