Skip to content
View adpth's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report adpth

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Adpth/README.md

Hi 👋 I’m Tharun Pasupuleti

Full-Stack Engineer · AI Engineer

M.S. Computer Science @ Purdue University Northwest B.Tech Computer Science & Engineering @ KITSW


Open to Full-Stack Engineer and AI Engineer roles in the US (on-site or hybrid).




💭 About Me

I build and ship full-stack products, and I work on the parts most people skip: concurrency, caching, failure handling, and making unpredictable systems behave predictably in production.

Two tracks I work in, and the projects below are sorted into both:

  • Full-Stack Engineering — React/Next.js and React Native on the front, Node.js/Express and FastAPI on the back, PostgreSQL and MongoDB underneath. Shipped to real users, including a government transit deployment live on Google Play.
  • AI Engineering — LLM systems built to be reliable, not demoed once: schema-enforced structured outputs, validation-and-retry loops, provider fallback chains, guardrails, and deterministic replay.

Always curious. Always building. Always learning.


💻 Full-Stack Engineering Projects

Ranked by production impact and system depth.


1. 🍔 Fudoo

Multi-vendor food delivery · Production launch

Hyper-local food ordering platform for college campuses, co-founded and shipped end to end.

  • Scaled to 500+ active users and 1,000+ weekly transactions at a 98%+ transaction success rate
  • Built an offline-first AsyncStorage queue with FIFO semaphore locking and idempotent transaction keys, so drivers keep working through campus network dead zones without duplicate state
  • Cut checkout queue time by ~50% using two-step Stripe authorization and capture, backed by a normalized Redux store for granular re-rendering


2. 🚍 TSRTC Medaram Jathara

Government deployment · Live on Google Play

Transit tracking app built with the Telangana State Road Transport Corporation for one of India's largest religious gatherings.

  • Served 15,000+ daily commuters across a four-day festival; 5k+ downloads on Google Play, officially recognized by the State Transport Corporation
  • Applied a 2dsphere geospatial index with MongoDB $near queries to return only buses and stops within a 5km radius, cutting payload size and client render overhead
  • Introduced Redis geospatial sorted sets between the API and MongoDB for sub-millisecond reads under peak load, with a background worker flushing telemetry to disk


➡️ View on Google Play



3. 🔗 ShortLnk

Full-stack SaaS · Frontend + backend repos

URL shortening platform with link management, custom redirect logic, and an analytics dashboard.

  • Served 1,000+ shortened links and 5,000+ clicks at sub-250ms redirect response times
  • Sustained 200+ concurrent requests under load testing without degradation
  • Held 80%+ test coverage across the API surface, cutting bugs by ~30% during development


➡️ Frontend · Backend

4. 💬 FeedbackFlow

Multi-tenant SaaS · Embeddable widget

Feedback collection platform teams drop into any site with a single script tag.

  • Improved API response times by ~25% through compound indexes on a multi-tenant MongoDB schema, and cut code duplication by 30%
  • Isolated the widget inside a Shadow DOM and excluded React from that build, keeping the embedded bundle under 30KB so host-site Core Web Vitals stay intact
  • Added dynamic per-domain CORS validation plus a Redis sliding-window rate limiter to block unauthorized embeds and traffic spikes



🤖 AI Engineering Projects

Ranked by engineering depth, not demo appeal. Every one treats the model as an unreliable dependency.


1. 🖥️ CUA System

Agentic automation · Deterministic replay

An LLM discovers how to complete a task in a hostile legacy web app once; that run is replayed in production with no model in the loop.

  • Records each discovery run as a typed, versioned artifact validated by a Zod schema, then replays it deterministically and returns a structured success / known-business-outcome / failure result
  • Perceives pages through the accessibility tree rather than screenshots or coordinates, resolving elements in an app with no test IDs, no real buttons, and iframe-nested content
  • Enforces a domain allowlist, risk classification, and non-retryable steps with human escalation on the live session; credentials never enter the model loop or the saved artifact


➡️ View Repo


2. 🧠 TaskPilot AI

AI SaaS · Structured generation

Turns a plain-text project goal into a validated, dependency-aware task roadmap.

  • Enforces a strict JSON schema in-prompt and validates with Pydantic, auto-retrying with the error trace appended so the client only ever receives structured data
  • Rejects circular dependencies through DAG validation before anything is committed to PostgreSQL
  • Traverses the full task graph using recursive CTEs in a single round trip, eliminating the N+1 problem and cutting read latency by 75% on large projects


➡️ View Repo


3. 🐦 TweetMind

LLM pipeline · Benchmarked and tested

Scrapes and clusters posts, detects trends, and drafts content into Notion for human review. It never auto-posts.

  • Measured 94.2% topic-cluster precision across 500 labeled tweets and 95.6% hook-pattern detection accuracy
  • Generates three drafts in 2.1s average / 3.4s P95, syncing 25 Notion pages in 0.82s at under $0.03 per full pipeline run
  • Ships log redaction, secret scanning in both pre-commit hooks and CI, and a hermetic pytest suite that runs with no live API credentials


➡️ View Repo


4. 🗂️ SprintForge

AI Kanban · Human-in-the-loop

Generative planning board that breaks project goals into prioritized, editable tasks.

  • Generated 35 tasks across 5 projects at 2.50s average roadmap generation with 100% JSON parse accuracy
  • Validates every AI response against typed interfaces server-side before insert, so malformed output never corrupts project state
  • Enforces organization-scoped RBAC in Next.js edge middleware, with optimistic UI that rolls back automatically when a mutation fails


➡️ View Repo


Also worth a look: X3D-VR for Aerospace (browser-based 3D learning platform at Purdue NW, 25% faster asset loading, 150+ students) · JSONLabs (frontend-only interactive JSON learning platform)


🧭 Experience

Role Organization When
Graduate Student Employee Purdue University Northwest Oct 2025 - Jan 2026
Software Engineer Accenture Feb 2024 - Aug 2024
Software Engineer Airbnb Mar 2021 - Jan 2024

🧰 Skills

💻 Frontend



React Native Redux Framer Motion Responsive Design

⚙️ Backend & APIs



REST APIs Microservices JWT / OAuth Rate Limiting WebSockets

🤖 AI Engineering



Structured LLM Output Prompt Engineering Provider Fallback Chains Agent Guardrails Gemini API

🗄️ Data & Cloud



Schema Design Query Optimization Caching Geospatial Indexing

🛠️ Engineering Practices



CI/CD Unit & Integration Testing Code Review Agile / Scrum Observability


📚 Currently Learning

  • Multi-agent orchestration and evaluation for LLM systems
  • Distributed system design for high-traffic real-time applications
  • AWS cloud architecture (working toward Developer Associate)

📫 Connect With Me

Website LinkedIn Email


✨ Always curious. Always building. Always learning. ✨

Pinned Loading

  1. TweetMind TweetMind Public

    An AI Twitter growth engine that scrapes viral posts, detects real-time trends, and drafts personalized tweets and replies, all synced to a Notion database for review.

    Python 1

  2. SprintForge SprintForge Public

    A modern, intelligent project planning tool that leverages Generative AI to automate the tedious process of breaking down high-level project goals into actionable, prioritized tasks.

    TypeScript

  3. TaskPilot-AI TaskPilot-AI Public

    TaskPilot AI is a premium, developer-centric AI Execution Operating System designed to eliminate procrastination, audit daily cognitive workload budgets, and decompose mental friction in real-time.

    TypeScript

  4. cua-system cua-system Public

    Computer-Use Automation (CUA) system with LLM-driven discovery and offline deterministic replay validation.

    TypeScript

  5. ShorkLnk-Backend ShorkLnk-Backend Public

    ShortLnk is a backend platform designed to efficiently transform long URLs into short, shareable links. It includes features like user registration, login, link creation, analytics tracking, and li…

    JavaScript

  6. ShortLnk ShortLnk Public

    ShortLnk is a full-stack URL shortening platform that allows users to create, manage, and analyse shortened links.

    JavaScript