# Arun Kumar Chukkala — Applied AI Engineer

> Arun Kumar Chukkala is an Applied AI Engineer in San Francisco who builds production LLM systems, agentic workflows, retrieval systems, evaluation harnesses, and AI guardrails.

- Location: San Francisco, CA
- Current role: Applied AI Engineer at IBM
- Availability: Open to Applied AI roles
- Email: [arunkiran721@gmail.com](mailto:arunkiran721@gmail.com)
- Canonical profile: [https://arun7.com/](https://arun7.com/)

## Selected work

### 01 — Cork

**Startup — AI Retail Operations Platform**

A production multi-tenant platform for independent stores — point-of-sale, inventory, purchasing, invoicing, payments, and loyalty in one system. Its AI layer turns paper distributor invoices into inventory updates, with human-in-the-loop confirmation so AI never writes to the books unverified.

- Invoice intelligence — Gemini OCR → structured extraction → five-stage matching cascade; human corrections become learned aliases, so matching improves without retraining.
- Evaluation-gated AI — an offline eval harness over real invoices holds the matcher at zero false positives, enforced as regression tests that block releases.
- Production discipline — per-tenant isolation verified by tests, idempotent financial writes, integer-cent money math, audit trails, and per-tenant AI cost caps.

Stack: TypeScript, React, Node.js, PostgreSQL, Prisma, pgvector, Gemini

Live project: [https://poscork.com](https://poscork.com)

### 02 — Claim Guardian

**AI Medical-Bill Auditor**

A free AI second opinion on any medical bill: reads a photo or PDF, prices every charge against real CMS Medicare rates, flags the overcharges, and drafts a cited dispute letter in ~60 seconds. Pricing is deterministic — the model explains findings but never invents a number.

Stack: FastAPI, Gemini, RAG, Next.js

Live project: [https://claim-guardian-v2-zyte.vercel.app](https://claim-guardian-v2-zyte.vercel.app)

### 03 — Agentception

**AI Job Search · LLM Evals**

Reads your resume, pulls live roles straight from applicant tracking systems, ranks them against you, and names the exact skills you are missing. Every claim is measured — an LLM eval harness over 69 real job descriptions (match AUROC 0.702, p = 0.017) that caught five bugs already shipped to production.

Stack: FastAPI, React, Voyage, Supabase

Live project: [https://agentceptionlanding1809.vercel.app/](https://agentceptionlanding1809.vercel.app/)

### 04 — Code Tribunal

**Multi-Agent Code Verification**

A multi-agent AI court that checks whether AI-generated code did what the ticket asked. Seven agents cross-examine every diff — hunting missing requirements and scope creep — and stamp a deterministic 0–100 trust score you can gate CI on. Ships as a web War Room, a CLI, and an MCP server.

Stack: Python, FastAPI, MCP, Next.js

Live project: [https://code-council.vercel.app](https://code-council.vercel.app)

### 05 — FinAgent

**Compliance-Grade Financial RAG**

A self-hosted research copilot for firms that ban public AI tools: six LangGraph agents run hybrid retrieval (BM25 + dense embeddings) over SEC filings, and a validator loops back on unsupported claims. Every answer is sealed into a SHA-256 hash-chained audit trail with an egress receipt naming exactly which providers touched the query — at ~$0.0005 per query.

Stack: Python, LangGraph, Qdrant, Next.js

Live project: [https://finagent-landing.vercel.app](https://finagent-landing.vercel.app)

## Production impact

- **62%** — Agent latency reduction — 850ms to 320ms via async execution + semantic caching
- **30%** — Manual review time cut with evaluation infrastructure and regression gates
- **98%+** — Uptime on production LLM services with retry, timeout, and fallback handling
- **0** — False positives on a 25-document evaluation harness, enforced by 15 regression tests

## Experience

### IBM — Applied AI Engineer

Mar 2024 — Present · United States

- Design and maintain evaluation infrastructure for LangGraph-based agent workflows in a regulated enterprise environment — golden datasets, regression gates, human-in-the-loop review — reducing manual review time by 30%.
- Reduced agent workflow latency by 62% (850ms → 320ms) through asynchronous execution and semantic caching of LLM calls.
- Ship Dockerized Python/FastAPI services orchestrating LLMs, internal tools, and enterprise data through CI/CD — 98%+ uptime with retry, timeout, and fallback handling around non-deterministic model calls.
- Build production observability that traces agent execution, tool calls, and failure modes end to end, converting production failures into permanent regression tests.

### Infosys — Software Engineer, Applied AI / Risk Platform

Jan 2021 — Dec 2022 · Hyderabad, India

- Engineered ensemble anomaly-detection workflows for fraud decisioning, reducing false positives by 18% on a regulated risk platform.
- Designed and deployed Dockerized FastAPI microservices and Azure Data Factory ETL pipelines processing 15M+ records.
- Implemented output-consistency monitoring that alerted on model drift before it reached downstream decisions.

## Education

- **M.S. Computer Science**, Lamar University — Dec 2024
- **B.Tech. Computer Science**, Sri Indu Institute of Engineering & Technology — May 2020

## Capabilities

### AI / LLM Systems

Agent loops & tool calling · Retrieval-Augmented Generation · Embeddings & vector search · Evaluation harnesses & golden datasets · Guardrails & grounded refusal · Human-in-the-loop design · Semantic caching · LLM observability & tracing · OCR / document intelligence · LangGraph

### Agentic Development

Claude Code · Cursor · Codex · Model Context Protocol (MCP) · Multi-agent code review · AI skills & internal tooling

### Engineering

Python · TypeScript · FastAPI · Node.js / Express · React · PostgreSQL / Prisma · Docker & CI/CD · Vercel / Render / Supabase · Sentry · Azure

## Research

### BARC: Blast-Radius-Aware Agentic Control

Preprint — 2026 / Agentic AI Safety

Agentic systems should be controlled by the consequence of their actions — not by model confidence. The paper introduces a tiered safety model for tools: read-only, draft, bounded write, and irreversible.

Paper: [https://arun7.com/papers/BARC_Paper_v2.pdf](https://arun7.com/papers/BARC_Paper_v2.pdf)

## Professional links

- [GitHub](https://github.com/arun3676)
- [LinkedIn](https://www.linkedin.com/in/arun-c-6703713ba/)
- [X](https://x.com/arunn0718)
- [Résumé](https://arun7.com/Arun_Chukkala_Resume.pdf)

Last updated: 2026-07-29