KarthikReddyYalala

AI Engineer, SDE, Tempe, ArizonaOpen to New Grad Software / AI Engineering roles, May 2027

I find the business problem, locate the AI opportunity, and ship the system end to end: architecture, APIs, front end, back end. Most of what I've built lately has AI inside it. The part I care about is making it reliable enough to hand to real users.

Karthik Reddy Yalala

Fortune 500production systems, shipped

4initiatives at Avis Budget

170+engineering hours saved

4.0GPA, ASU '27

Avis Budget Group

AI Enablement Intern, Fortune 500, Parsippany, NJ, May to Aug 2026

engineering & product hours saved
170+
projected annual cost savings
$1,000,000+
initiatives shipped
4

Embedded with the Customer Experience team on a CEO-sponsored AI initiative. Four end-to-end initiatives, from system architecture and API design through front-end and back-end implementation to production handoff. Designs presented to the Director and VP of Customer Experience; the platform demoed to the CEO.

Internal AI assistant

  • Bedrock AgentCore
  • Azure AD SSO/MFA
  • SSE streaming
  • DynamoDB

Gating publish behind a test felt like friction until you imagine the alternative: an untested prompt going live to the whole company.

An employee-facing assistant for company policy and compliance questions, built on AWS with AWS AgentCore at the core. AgentCore is deliberately constrained: retrieval pulls the top-ranked chunks from the Knowledge Base based on the live config and passes them to AgentCore, which wraps the Claude call, enforces guardrails, handles multi-turn session memory, and formats the response. Admins manage everything through a web console. Uploads go straight to S3, a Lambda triggers KB resync, and new content is live without a deploy. Employees log in with SSO and chat. Admins get the pipeline. Users get the interface.

  1. Draftadmin edits prompt

    safety scanleads toreject

  2. Testagainst the draft, live

  3. Approvecannot skip testing

  4. Publishre-verified, versioned

    zero downtime

  • Passed a 6-phase test suite: 100% faithfulness, 0 hallucinations, 5/5 injection attempts blocked
  • Constrained the model with AgentCore: guardrails, memory, and formatting on every call
  • Shipped a zero-downtime config pipeline: draft, test, approve, publish, versioned
  • Delivered real-time streaming answers with multi-turn session memory
  • Replaced a third-party LLM SaaS dependency with an in-house AWS stack, keeping employee and compliance data in the company's own environment
Read the engineering decisions

Why a system prompt is treated as executable code

It is attack surface. Anything an admin writes passes two independent scans before it is stored: a pattern denylist for known jailbreak shapes, and a model-based review for what the patterns miss. If either objects, nothing is written at all.

Why you cannot approve what you have not tested

The console gates each step on the previous one: save, then test against the draft, then approve, then publish. Approval is re-verified at publish time, and editing after approval resets the gate. Config changes go live without a restart, and if the config store is unreachable the service falls back to a known-good default with a visible warning instead of failing.

Why follow-up questions get rewritten before search

"How do those two relate?" is meaningless to a search index. Before retrieval, a small fast model resolves pronouns and references against the recent conversation into a standalone question. It is skipped entirely on the first turn, so the common case costs nothing.

Natural-language-to-SQL analytics engine

  • Amazon Bedrock
  • Claude Sonnet
  • Vercel AI SDK
  • Tool-use agent
  • Next.js
  • PGlite
  • SQL

Seven column-name mismatches. Mock data would have let every one of them ship, and each would have failed in front of a user.

A self-correcting agent loop that turns a plain-English question into a table, a summary, and a chart. No SQL, no data-team ticket, no waiting a day for someone to pull it. It runs as a tool-use loop on the Vercel AI SDK over Bedrock Claude: the agent plans its approach, pulls only the schema it needs, writes the SQL, guards and executes it, then presents the result. If a query fails, it reads the error and corrects itself. I analyzed the full database and scoped the agent to the business-relevant tables only, which cut token cost and latency sharply with nothing lost in coverage. Verified against a gold question set before anyone relied on it. Customer PII is structurally incapable of reaching the output.

  1. Questionplain English

  2. Planagent picks the approach

  3. Schemaonly the tables needed

    PII columns excluded

  4. SQLwritten, self-corrects on error

  5. ExecuteSELECT-only, PII-blocked

    blockedleads torewrite

  6. Resulttable + summary + chart

  • Reduced data turnaround from days to seconds
  • Validated against a gold-question evaluation set covering every business domain
  • Scoped the schema to business tables only for lower token cost and faster plans, with no coverage lost
  • Shipped a KPI dashboard so the team reads regular metrics without a query
  • Saved 50 engineering and 20 product hours
  • Handed off production-ready with docs, schema, and runbook
Read the engineering decisions

Why it's an agent loop

One prompt writing SQL in a single shot fails silently the moment a column name is off or a join doesn't hold. Instead the model works as an agent with a small set of tools (plan, read schema, execute, present) and a hard stop once an answer is produced. Every query passes a guard before it touches the database: SELECT-only, no restricted columns, no PII. If execution fails, the error goes back to the agent and it tries again. That loop is what makes it reliable enough to hand to a non-technical team.

Why the agent sees only a scoped schema

I analyzed the entire database and found most of it was operational tables the questions never needed. Handing all of that to the model would have meant more tokens, slower plans, and more ways to write a wrong join. So I scoped the agent to the business-relevant tables only. That meant smaller prompts, lower cost, faster answers, and no loss of coverage on the questions people actually ask.

Why PII is removed from the schema, not filtered from the output

A filter is something you can forget to apply. Sensitive customer fields were excluded from the schema the model can see at all, and the remaining sensitive fields were restricted to aggregate use only. The model cannot select a column it was never shown.

AI delivery assistant

  • Rovo AI workflows
  • Jira automation
  • Scheduled triggers

The least sophisticated thing I built and one of the most used. A schedule, a query, an email.

Delivery tracking that reports itself. Engineers get a personalized morning digest of what is overdue, due today, and coming up; managers get a weekly board-health snapshot. Deployed organization-wide.

  1. Scheduled scandaily + weekly

  2. Group by ownerper-assignee lookup

  3. Categorizeoverdue / today / upcoming

  4. Deliverpersonalized digest

    live dashboard

  • Deployed org-wide, automating the daily overdue-ticket triage a delivery lead did by hand
  • Bypassed Jira's 100-issue automation lookup cap with REST web requests for full board coverage
Read the engineering decisions

Why the report goes to the inbox

A dashboard only helps the people who remember to open it. Routing each engineer only their own items, in their inbox, before the day starts, is what actually changed behaviour. The live dashboard is there for drill-down.

AI visibility research

Beyond the build

Delivered to the VP of Customer Experience

Alongside the engineering work, I joined the weekly Customer Experience meetings. I took the initiative to investigate how AI is reshaping the way customers choose a car rental provider. I ran a competitive analysis across the company and its competitors using public customer data, and delivered a written report with findings and recommendations. I also thought a real-time view of public customer sentiment would be useful for the team, so I built a prototype social listening dashboard that aggregates reviews across Google, Trustpilot, the App Store, Google Play, and Reddit, surfacing sentiment trends and billing dispute patterns in one place.

View the prototype dashboard

Things I built to find out if they'd work.

Agent systems, mostly. A few are experiments I kept because of what they taught me.

Crucible

2026

An AI mock interviewer that pushes back on vague answers and remembers where you struggled.

Five LangGraph agents run each interview: intake reads your resume and the job description, the planner builds the question plan, Marcus the interviewer asks one follow-up at a time, the evaluator decides whether each answer would survive a real interviewer, and memory carries your weak spots into the next session. I measured it on 76 hand-labeled interview situations, each run three times: 76/76 correct decisions, 34/34 correct verdicts, and prompt injection never changed a decision. That eval caught an SDK bug that would have broken every live model call while 1,500+ unit tests passed. It also has a live coding round with Python running in your browser across all 75 Blind 75 problems, streaming voice, and an optional real video avatar.

  • LangGraph
  • Amazon Bedrock
  • FastAPI
  • React
  • DynamoDB
  • AWS Lambda
  • Cognito
  • Pyodide
  • Deepgram
  • ElevenLabs
  • Tavus

Open AutoData

2026

Hack Day 2026

Five local-model agents that write, filter, and grade fine-tuning datasets.

Built at Hack Day 2026 on Meta FAIR's Agentic Self-Instruct method: agents write questions from documents, check for leakage, and grade answers from a weak (Llama 3.2 3B) and a strong (DeepSeek V4.1 Flash) model. A question is kept only when the weak model scores 65% or below, the strong model scores 60-95%, and the gap exceeds 20 points. Against Meta's paper: a 56.5-point weak/strong gap vs. 31.4, and an 84.2% vs. 77.2% strong-solver average.

  • LangGraph
  • OpenRouter
  • Llama 3.2
  • DeepSeek
  • Nemotron
  • Snowflake
  • Next.js
  • Tailwind
  • pytest

MCP Knowledge Base

2026

Private docs inside Claude, with no uploads, no cloud, and exact citations.

An MCP server that indexes your markdown files locally and gives Claude search through two tools, search_docs and list_docs, with a line-range citation on every answer. It combines embeddings with BM25 keyword search, then a cross-encoder reranks the top 20; only the best 5 chunks enter the conversation. Benchmarked on BEIR with significance tests: nDCG@10 of 0.720 on SciFact (+10% over v1) and 0.364 on NFCorpus (+18%), matching qmd on SciFact at 67 ms per search versus 7 s.

  • MCP
  • Python
  • ChromaDB
  • BM25 (SQLite FTS5)
  • all-MiniLM-L6-v2
  • Cross-encoder reranking
  • Ollama

Deep Research Agent Team

2025

A planner, executor, and validator loop that does the reading for you.

A modular multi-agent research system where a planner decomposes a topic, executors gather sources, and a validator checks the synthesis before it is returned. Built with retry logic, exponential backoff, and output caching so a single failed tool call does not sink an entire run. Cut manual research effort by 70%.

  • CrewAI
  • LangGraph
  • OpenAI APIs
  • Pinecone
  • Python

FinRAG MCP

2026

Cited financial intelligence over 500+ SEC filings, served through an MCP server.

An MCP server that answers financial questions with source citations across 500+ SEC filings from 50 companies, refreshed weekly from EDGAR with table-aware extraction. Retrieval runs four stages: Claude Haiku query rewriting, hybrid BM25 and dense search on Bedrock Knowledge Bases, neural reranking, and a numerical grounding verifier that catches hallucinated figures. A 300-question FinanceBench set with ragas metrics runs as a regression gate in GitHub Actions.

  • Amazon Bedrock
  • MCP SDK
  • Python
  • FastAPI
  • ragas
  • AWS Lambda
  • DynamoDB
  • CloudWatch

Autonomous Trading Floor

2025

Four agents that argue about a trade before making it.

A simulated trading desk where an analyst, a trader, a risk manager, and a critic evaluate the same position and have to reach a decision. It's wired across six MCP servers exposing 44 tools, so the hard part turned out to be tool routing.

  • MCP
  • CrewAI
  • Docker
  • OpenAI APIs
  • Python

Chatify

2024

Real-time messaging, built from the socket up.

Chatify screenshot

A full-stack messaging application with bi-directional event-based communication over Socket.io, on a Node and Express backend. I built it to learn connection lifecycle, room state, and reconnection handling.

  • Node.js
  • Express
  • Socket.io
  • MongoDB

Stock Price Predictor

2024

An LSTM that beat the linear baseline by 12%.

A sequence model for price forecasting over 50k+ data points, with normalization, windowing, and hyperparameter tuning. The useful lesson was in the evaluation: a model that looks good on a chart can still be worse than the baseline you didn't bother to run.

  • Python
  • TensorFlow
  • Pandas
  • Matplotlib

What I actually use.

Grouped by what it's for. No proficiency bars, because a number I assign to my own skill isn't evidence of anything.

Software Engineering

How the work gets done

  • System architecture
  • REST API design
  • Unit testing
  • Feature testing
  • Code review
  • CI/CD
  • Agile
  • Git / GitHub

AI / GenAI

What I reach for first

  • Amazon Bedrock
  • Bedrock AgentCore
  • Bedrock Knowledge Bases
  • Bedrock Guardrails
  • Claude
  • OpenAI APIs
  • LangGraph
  • LangChain
  • CrewAI
  • RAG
  • RAG evaluation
  • Agent loops / tool use
  • Vector search
  • Embeddings
  • Prompt engineering
  • PyTorch
  • TensorFlow
  • Hugging Face

Languages

  • Python
  • TypeScript
  • JavaScript
  • C/C++
  • Java
  • SQL
  • HTML/CSS

Web & APIs

How it reaches a user

  • React
  • Next.js
  • Node.js
  • FastAPI
  • Express
  • REST
  • WebSocket
  • SSE
  • Tailwind CSS

Cloud (AWS)

Where it runs

  • Lambda
  • API Gateway
  • CloudFront
  • S3
  • DynamoDB
  • Cognito
  • CloudWatch
  • CloudTrail
  • WAF
  • Azure AD
  • Docker

Databases

  • PostgreSQL
  • MySQL
  • MongoDB
  • DynamoDB
  • Redis
  • Firebase

How I got here.

Sep 2026 to Apr 2027

AI Systems Architect & Guardrails Lead

ProfessorDM Technologies Inc., ASU Capstone, Remote (New Brunswick, CA)

  • Leading technical architecture for a seven-person capstone team building a production AI SaaS platform that converts Instagram DM traffic into monetized, searchable expert knowledge for a creator with 165K+ followers.
  • Owning cloud infrastructure, CI/CD pipelines, data architecture, authentication/RBAC, secrets management, observability and release gates across five integrated technical workstreams.
  • Co-designing AI guardrails, prompt/model change controls and evaluation gates to ensure no AI-generated response reaches consumers without creator review and multi-stage safety filtering.
  • Defining interface contracts, security standards, and architecture decision records that five workstream leads build against, coordinating cross-service integration from Instagram webhooks through AI classification, semantic retrieval, creator review and answer delivery.
  • Accountable for security hardening, regression testing strategy, performance reviews and operational readiness ahead of controlled live trials with real users on the Sweetlife Flora Instagram account.
  • AWS
  • Supabase
  • pgvector
  • Vercel
  • Stripe
  • Meta Graph API
  • CI/CD
  • RBAC
  • Observability

May to Aug 2026

AI Enablement Intern

Avis Budget Group, Parsippany, NJ

  • Spearheaded four end-to-end software initiatives from system architecture and API design through front-end and back-end implementation, saving 170+ engineering and product hours.
  • Presented designs to the Director and VP of Customer Experience; demoed the platform directly to the CEO.
  • Delivered AI visibility research to the VP of Customer Experience.
Read the case study
  • Amazon Bedrock
  • AgentCore
  • Guardrails
  • AWS Lambda
  • API Gateway
  • DynamoDB
  • Azure AD

2025

AWS Certified Cloud Practitioner

Amazon Web Services, CLF-C02

May to Jul 2025

Data Analyst Intern

Food Forest AI, Philadelphia, PA

  • Engineered GPT-powered extraction pipelines to scrape, parse, and validate 200+ company records against agreed schemas, improving dataset quality by 40%.
  • Automated deduplication and anomaly flagging, removing manual cleaning from the ingestion path.
  • Python
  • OpenAI APIs
  • Web scraping
  • Data validation

May to Jul 2024

Software Engineer Intern

YJR Realtors, Hyderabad, India

  • Built and deployed a React/Node.js platform with a conversational AI agent using LangChain, OpenAI API, and Pinecone vector search for natural-language listing queries and 24/7 tour scheduling.
  • Engineered an AWS Textract document processing pipeline to parse deeds, leases, and contracts, extract metadata, and flag compliance risks and financial exposures before closing.
  • Automated backend workflows using S3 and SendGrid to streamline contract ingestion, secure document storage, and customer communications, improving team response time by 40%.
  • React
  • Node.js
  • LangChain
  • OpenAI API
  • Pinecone
  • AWS Textract
  • S3
  • SendGrid

2023 to 2027

B.S. Computer Science

Arizona State University, Tempe, AZ

  • 4.0 GPA. Dean's List every semester: Fall 2023 through Fall 2025.
  • Coursework: Algorithms, Data Structures, Operating Systems, Software Engineering, Artificial Intelligence, Machine Learning, Cybersecurity, Discrete Mathematics.

About

Karthik Reddy Yalala

I'm a senior at Arizona State studying computer science. Most of what I know about building software I learned by shipping things other people then had to rely on.

This summer at Avis Budget Group I took four initiatives from a blank architecture diagram to production handoff: a multilingual retrieval service, an internal assistant, a natural-language analytics engine, and a delivery automation. I owned the system design, the API surface, and both ends of the stack. I presented the designs to a Director and a VP, and demoed the platform to the CEO.

The part I care about most is making sure the thing actually works before anyone else has to rely on it. End-to-end testing, finding the edge cases before users do, and making sure the code is solid enough to ship. That's usually where the difference between a demo and a product lives.

CurrentlySeptember 2026

Back at ASU for senior year. Leading the technical architecture for ProfessorDM, a seven-person capstone building a production AI platform against a 165K-follower Instagram account. Owning cloud, data, security, AI guardrails and cross-workstream integration.

Let's talk about what you're building.

Open to New Grad Software / AI Engineering roles, May 2027

↑↓ move↵ openesc close