Apoorav Rao.
Available · AI Agent Developer

I build agents
that actually ship.

I'm Apoorav Rao — CS final-year, graduating June 2026. I design production-grade agentic systems with multi-step tool-calling, real failure handling, and human-in-the-loop where it matters.

Currently building TrainFlow — AI fitness coaching for iOS, serverless AWS backend.

● Open to remote·Python · AWS · LangChain · OpenAI·4+ shipped agents
View projectsDownload CV ↓Get in touch
scroll
01 ·

About

CS student graduating June 2026, building at the intersection of AI agents, cloud infrastructure, and native mobile. My work is production-grade — not demos.

Currently building TrainFlow, an AI fitness coaching platform for iOS with HealthKit integration and a fully serverless AWS backend.

AI & LLMs

Building agents, knowledge bases, and LLM-powered workflows. From evaluation frameworks to autonomous branding agents.

Data Engineering

Designing ETL pipelines with Apache Airflow, dbt, and PostgreSQL. Making raw data actionable at scale.

Cloud Infrastructure

Serverless-first with AWS Lambda, DynamoDB, API Gateway, and CDK. Infrastructure as code, always.

iOS Development

Native SwiftUI apps with HealthKit, WatchKit, and deep Apple ecosystem integration.

Timeline

2022

Started CS degree

Fell in love with algorithms

2023

First AWS projects

Lambda, DynamoDB, serverless

2024

LLM integrations

Agents, RAG, knowledge bases

2025

Internship @ Caterpillar

LangChain · Airflow · AWS

2026

Graduating

Ready for what's next

02 ·

Work

one role, real numbers
Jun 2024 — Mar 2025Internship · 10 months

Data Engineering Intern

Caterpillar Signs Pvt. Ltd. (Group Bayport) · Gurugram
  • ▸Built an LLM-powered data pipeline ingesting CRM data from Freshdesk and Salesforce APIs, running zero-shot classification and extractive QA on support tickets via LangChain, then transforming and loading into PostgreSQL (AWS RDS) via dbt.
  • ▸Orchestrated the entire pipeline with Apache Airflow — handling retries, backfill, and dependency management across daily and weekly jobs.
  • ▸Applied multi-label zero-shot classification to categorise tickets without a labelled training set, surfacing 20%+ more hidden issues buried in unstructured text.
  • ▸Reduced manual ticket review time by 400% by replacing a human-review workflow with an automated LLM audit pipeline with human-in-the-loop escalation.
LangChainApache AirflowdbtPostgreSQLAWS RDSFreshdesk APISalesforce APIPython
400%ticket auditing efficiency15%reduction in ticket volume20%+more issues surfaced
Education
🎓
BRCM College of Engineering and Technology
B.Tech in Computer Science
Bahal, Haryana · 2022 – 2026
📚
O.P. Jindal Modern School
12th CBSE
Hisar, Haryana · 2022
A couple of wins
🥈Problem-solving, coding, and innovation track.2nd Prize — Collegiate Hackathon
🏆National hackathon — selected in the top 10 teams out of hundreds of applicants.Top-10 Finalist — IIM Ahmedabad × Ashoka University
04 ·

Selected agents

code, not concepts
423kcalCalories
5.2kmDistance
142bpmHeart Rate

TrainFlow

A production-grade iOS application that ingests 20+ Apple HealthKit biometrics daily and serves them to an agentic AI coaching loop. A context builder pre-loads user health and training snapshots, a tool executor dispatches up to 5 OpenAI function-calling rounds per request, and a secondary model generates plain-English health summaries. User identities are extracted from Cognito JWT claims — never request bodies.

  • ▸24 Lambda functions + 6 DynamoDB tables deployed via CDK
  • ▸Agentic loop: up to 5 OpenAI function-calling rounds per coaching request
  • ▸Apple Watch companion syncing biometrics in real time via HealthKit
  • ▸Cognito JWT authentication — zero PII in request bodies
SwiftAWS LambdaDynamoDBCDKOpenAIHealthKitCognito
KBArticleArticleArticleArticleTagTag

LLM Knowledge Base

A shell-powered pipeline inspired by Andrej Karpathy's knowledge management approach. It ingests raw articles, runs LLM summarisation and keyword tagging on each one, then stitches the outputs into an interconnected wiki with cross-linked entries. The result is a fully queryable, self-updating knowledge graph that grows every time you drop a new document in.

  • ▸Zero-configuration ingestion — drop any article into the watch folder
  • ▸LLM-powered summarisation, tagging, and cross-reference extraction
  • ▸Outputs a browsable static wiki with bidirectional links
  • ▸Inspired by Karpathy's personal knowledge management patterns
ShellPythonLLMKnowledge GraphKarpathy
api-agent — running tests
$ test-agent parse openapi.yaml_
Generating & running tests0%

API Testing AI Agent

An AI agent that reads OpenAPI/Swagger specifications and autonomously generates comprehensive test suites covering happy paths, edge cases, auth failures, and schema validation errors. It executes the generated tests against live endpoints and produces a structured report with per-endpoint pass rates and failure diagnostics — no manual test writing required.

  • ▸Parses OpenAPI 3.x and Swagger 2.0 specs automatically
  • ▸Generates tests for happy path, error codes, edge cases, and auth failures
  • ▸Executes tests against live endpoints and captures response diffs
  • ▸Produces structured pass/fail report with failure diagnostics
PythonOpenAIAPI TestingOpenAPIAutomation
ProfileAgentGitHubinLinkedIn↓Resume✓ Synced automatically

Profile Agent

An autonomous agent that monitors GitHub for new repositories, analyzes each project's README, languages, and screenshots, and generates a LinkedIn post, edits a LaTeX resume, and updates the GitHub profile README — all in a single GPT-4o call. The agent reads the full history of past posts before generating anything so tone rotates and content never repeats. Runs daily via macOS launchd; the resume compiles to a one-page PDF via Tectonic with automatic project rotation.

  • ▸Monitors GitHub via webhooks and daily polling — triggers on new repos
  • ▸Single GPT-4o call generates LinkedIn post + resume edit + README update
  • ▸Reads full post history before writing — ensures non-repeating, rotating tone
  • ▸Resume auto-compiles to PDF via Tectonic with one-page constraint enforced
PythonGPT-4oGitHub APILinkedIn APILaTeXSQLitelaunchd
Model AVSModel B
Model A
Model B

AI-Duel

A model evaluation framework that fans the same prompt out to multiple LLMs in parallel, renders their outputs side-by-side with token counts and latency, and applies a configurable scoring rubric. Designed to make switching between model providers fast and evidence-based — see exactly where GPT-4o beats Claude or vice versa on your specific workload.

  • ▸Parallel inference — sends one prompt to N models simultaneously
  • ▸Side-by-side diff view with latency, token count, and cost estimates
  • ▸Configurable scoring rubric: factuality, tone, length, task completion
  • ▸Export results as JSON for downstream analysis or CI/CD integration
TypeScriptLLMEvaluationMulti-modelBenchmarking
✌️V
→
MediaPipe
Hello
ASL → English

Sign Language Translator

A computer vision pipeline that captures hand landmarks in real time using MediaPipe, normalises the 21-point hand skeleton to a pose-invariant feature vector, and feeds it into a custom-trained neural network that classifies ASL gestures with high accuracy. The system runs at 30fps on a laptop CPU and outputs translated text with confidence scores displayed live on screen.

  • ▸21-keypoint hand skeleton extracted via MediaPipe at 30fps
  • ▸Custom neural network trained on a normalised ASL dataset
  • ▸Pose-invariant feature extraction — works regardless of hand size or distance
  • ▸Live on-screen overlay showing gesture class and confidence score
PythonMediaPipeDeep LearningOpenCVComputer Vision
AAPL+2.4%
$10,423.20
BUY

Paper Trading Web App

A full-stack paper trading platform that simulates real market conditions. Users start with a virtual portfolio and can place market/limit orders, track P&L over time, and review candlestick charts with technical indicators. The Django backend handles order matching, portfolio accounting, and historical data storage — making it a realistic environment for developing and backtesting trading strategies.

  • ▸Django REST backend with real-time order matching and portfolio accounting
  • ▸Candlestick charts with moving averages and volume indicators
  • ▸Supports market orders, limit orders, and stop-loss triggers
  • ▸Full trade history with P&L analysis per position
PythonDjangoMySQLREST APIBootstrap
03 ·

The stack

things I reach for

Agentic AI

OpenAI Function CallingLangChainMulti-step Tool ExecutionAgentic LoopsKimi K2Prompt Engineering

Cloud & Backend

AWS LambdaDynamoDBAPI GatewayCDKCognitoRDSS3PythonREST APIs

Data & Pipelines

ETL / ELTApache AirflowdbtPostgreSQLSQLitePandas

iOS & Mobile

SwiftUIHealthKitWatchKitCoreData

Languages

PythonSwiftTypeScriptSQLShell

Tools & APIs

GitGitHub APILinkedIn APINext.jsReactTailwind CSS
Exploring
AWS BedrockMulti-agent systemsRAG pipelinesHealthKit MLVector DBs
05 ·

Let's talk

i reply fast