Raghav Anand

Portfolio

AI prototypes for regulated work

Hands-on builds where rules handle the exact work and AI handles the messy leftovers. All data is synthetic; all code is public.

Healthcare · Agents

Claims Pre-Check Agent & Rx Adjudication API

PythonClaude APIEvalspytest
View on GitHub →
Problem

Pharmacy claims get denied for issues that could have been caught before submission, and some of those issues hide in free-text notes that rule-based checks can't read.

What I built

Deterministic rule checks for the exact issues, a Claude-based agent that reviews the notes, and a pharmacy adjudication API to test against.

Outcome

On a labeled synthetic eval set, the agent caught risks in free-text notes that the rules alone missed. The adjudication API ran at ~1.3 ms p95 in local load testing.

Fintech · Data

3-Way Reconciliation Data Generator

PythonpandaspytestFintech
View on GitHub →
Problem

Reconciliation automation is hard to judge without realistic data and a known right answer for every mismatch.

What I built

A generator for ledger, processor and sponsor-bank records covering cash advances and repayments, with injected breaks: ACH returns, late postings, duplicates and fee mismatches.

Outcome

Every injected break ships with an answer key, so any matching approach, rules or AI, can be scored instead of eyeballed.

Agents · Fundamentals

Tool-Use Agent

PythonClaude APITool use
View on GitHub →
Problem

Agent frameworks hide the core loop, which makes it hard to reason about cost, failure and control.

What I built

A minimal agent that decides when to call a Python tool, reads a local file and answers from what it found.

Outcome

The request, execute, return loop behind every AI agent, in a few readable files.

LLM Ops · Guardrails

Text Summarizer CLI

PythonClaude APIGuardrails
View on GitHub →
Problem

LLM calls can fail in quiet ways or cost more than expected once real inputs arrive.

What I built

A command-line summarizer with input limits, error handling, and a cost estimate shown before each call.

Outcome

Each estimate is checked against actual token usage, so the guardrail is measured rather than assumed.

See everything on GitHub →