DEV Community: DEVANSHU PURI

BaseBuildStack: Privacy-First AI Technical Interview Preparation Built on Open Weights

DEVANSHU PURI — Sun, 18 Oct 2026 18:49:32 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

BaseBuildStack is an engineering interview preparation micro-SaaS powered by open-weight AI models and deterministic validation engines.

Traditional technical interview preparation is broken in two major ways:

Exorbitant Cost: Commercial platforms charge candidates anywhere from $35 to $100+ per month, pricing out students, career switchers, and engineers from emerging economies.
Data Privacy Risks: Closed platforms upload sensitive candidate resumes, GitHub repositories, and private code solutions to third-party commercial LLM providers to train proprietary models.

BaseBuildStack delivers an adaptive, high-speed interview practice suite that runs on open-source AI:

Module I: DSA & Technical Blitz: 45-second pattern identification cards, line-spotting bug audits, and SM-2 spaced repetition review.
Module II: Personalized Interview Quizzes: Adaptive question drills generated from the candidate's exact tech stack, resume claims, and code repositories.
Module III: Tech Bites: 15-minute interactive architectural conceptual models (RAG pipelines, Cache-Aside strategies, Vector Embeddings, Docker process isolation).
Module IV: Mind Sprints: Daily 5-minute single-solution logic grids and 4-quadrant metacognitive confidence calibration.

Who Would Pay For It?

Job-seeking software engineers, bootcamp graduates, and developers preparing for AI and distributed systems technical rounds. We designed it as a sustainable side hustle with a ₹9/month (~$0.11 USD) Pro Plan via Razorpay, unlocking deep repository analysis, adaptive weak-area drills, and unlimited review sessions.

Demo

🌐 Live Deployed App: https://basebuildstack.onrender.com
🚀 Test Account Available: Full guest mode or instant registration with Google OAuth, GitHub OAuth, or Email/Password.

Code

devanshu-puri / BaseBuildStack

This is a Next.js project bootstrapped with create-next-app.

Getting Started

First, run the development server:

npm run dev
# or
yarn dev
# or
pnpm dev
# or
bun dev

Open http://localhost:3000 with your browser to see the result.

You can start editing the page by modifying app/page.tsx. The page auto-updates as you edit the file.

Learn More

To learn more about Next.js, take a look at the following resources:

Next.js Documentation - learn about Next.js features and API.
Learn Next.js - an interactive Next.js tutorial.

You can check out the Next.js GitHub repository - your feedback and contributions are welcome!

Deploy on Vercel

The easiest way to deploy your Next.js app is to use the Vercel Platform from the creators of Next.js.

Check out our Next.js deployment documentation for more details.

View on GitHub

The complete source code is open source under the MIT License on GitHub:

👉 https://github.com/devanshu-puri/BaseBuildStack

How I Built It

BaseBuildStack was architected from day one to treat open-source AI as a first-class citizen rather than a bolt-on API wrapper:

Local Open-Weight Inference (Ollama): Our core LLM Gateway interfaces with local open-weight models (gemma2:9b and llama3) via Ollama. It synthesizes targeted interview questions, hints, and explanations on demand.
Open Embeddings (all-minilm): We utilize open SentenceTransformer embeddings via Ollama for repository chunking, question similarity search, and semantic tag matching.
Deterministic Assertion Sandbox: Open models are powerful, but non-deterministic hallucinations break technical learning. We pair all model code evaluation with an isolated Node.js worker-thread sandbox (tests/unit/sandbox.test.ts) that executes generated test suites deterministically against candidate code.
Zod Boundary Schemas: Every response generated by open-source models is parsed and strictly validated against Zod schemas (LlmValidationError) to guarantee rigid runtime types.
Modern Web Stack: Next.js 15, React 19, custom Aurora glassmorphism UI system, user-scoped MongoDB storage, and Web Crypto PBKDF2 authentication.

Why Does Open Innovation Matter?

Open innovation was not just a design preference for BaseBuildStack—it is the foundational reason the product works:

Unit Economics That Make Micropayments Viable: If we used closed, commercial APIs charging $0.01 to $0.03 per prompt/completion, a candidate completing 50 drills a day would burn through our margins in hours. Running open-weight models on self-hosted or local infrastructure reduces marginal inference costs to near zero, allowing us to offer a profitable ₹9/month subscription model.
Zero-Trust Resume & Code Privacy: Candidates often practice with proprietary codebases or sensitive career history. By offering a local inference pipeline via Ollama, code and resumes remain strictly on the candidate's machine and are never shipped off to train third-party commercial datasets.
Model Autonomy & Future-Proofing: Closed API providers frequently deprecate models, shift safety guardrails, or unexpectedly raise pricing tiers. With our open gateway architecture, we can hot-swap between Gemma, Mistral, Llama, or DeepSeek Coder with a single environment flag without changing any application logic.
Offline Resilience: Because the core engine, spaced-repetition scheduler, and logic puzzle generator are fully deterministic, the app can function completely offline on a laptop without internet connectivity.

My Agent Session

This entire platform was engineered, architected, and continuously validated using the Google DeepMind Antigravity AI pair programming agent and DevRelay.

From writing the OWASP-grade cryptographic authentication to resolving strict Next.js 15 middleware boundaries and validating all 251 unit test suites, the transparent workflow was recorded via DevRelay.

My First User

We onboarded two candidate testers actively interviewing for full-stack and platform roles.

Tester 1 (Student): Specifically praised the 45-second Pattern Blitz cards, noting that fast-paced pattern recognition helped reduce anxiety during real live-coding rounds.
Tester 2 (Junior Developer): Tested the resume-tailored quiz generation and appreciated that the app didn't require sharing API keys or uploading confidential work projects to third-party closed servers.
Payment Feedback: Both confirmed that ₹9/month was a micro-expense they would pay without hesitation compared to costly recurring subscription models.

Prize Categories

Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

OutLoot System Architecture

DEVANSHU PURI — Sat, 10 Oct 2026 17:58:24 +0000

OutLoot: Turn Screen Time Into Real Loot with Local Gemma AI & Verified Outdoor Trails

DEVANSHU PURI — Sat, 10 Oct 2026 17:46:11 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Most fitness or location-based games commit one of two fundamental mistakes:

They demand screen addiction: They force you to stare directly at your glass display while navigating busy sidewalks and nature paths, turning an outdoor walk into another session of visual fatigue.
They are trivially spoofed: They rely on easily faked client timers or step counts that can be simulated by shaking a phone from a desk chair or riding in a car.

OutLoot flips this entirely. It is a screen-light, server-authoritative exploration game designed for developers, students, and knowledge workers who spend 8+ hours a day staring at code editors and need a compelling, gameful reason to step away, disconnect, and touch grass.

Key Highlights:

Server-Authoritative Anti-Spoofing Movement Engine: The client collects raw GPS breadcrumbs, but never validates its own progress. The server evaluates coordinates against physics constraints (speed between 0.5–3.5 m/s, GPS horizontal accuracy ≤ 30m, sample interval ≤ 30s, and a minimum displacement threshold of 75m) to ensure only genuine outdoor foot travel earns rewards.
Local Gemma AI Storytelling & Quests: A local open-weight Gemma model (gemma2:2b via Ollama) serves as the game master. It generates dynamic outdoor scavenger quests, walking tips, and mythic relic lore (like Clover Sprouts, Dewdrop Crystals, and Ancient Acorns) without sending any location data to third-party cloud APIs.
Active Screen Time Heartbeats & Grass Debt: A background engagement tracker monitors active mouse, keyboard, and touch interactions. When you exceed your threshold (e.g., 30 minutes of continuous screen time), a frosted-glass Touch Grass Alert modal nudges you to take a walk. You can also import daily screen hours from Google Digital Wellbeing to calculate your Grass Debt (15 outdoor minutes required per hour of screen time).
Social Leaderboard, Follow System & Trail Inspector: Compete on a global leaderboard ranked by Grass Score (points + outdoor_mins * 2). Follow teammates, send tactile 🌱 Nudge alerts to get friends outside, and click 🗺️ Trail on any player to inspect their verified outdoor walking path directly on the retro dark-mode map!
Double-Entry Points Ledger & Real Perks: All points are recorded in an append-only ledger with cryptographic idempotency keys. Redeem points for promotional partner vouchers with unique redemption codes.
Zero Commercial API Keys & 100% Open Data: Built with OpenStreetMap (OSM) and Leaflet using custom CSS dark styling. No Mapbox/Google Maps billing, no watermarks, and no third-party telemetry.

Demo

Live Deployed App (Render): https://outloot-tgtrails.onrender.com
Local Run: Run uvicorn server.main:app --reload and visit http://127.0.0.1:8000/.
Zero-Setup PWA: Fully responsive across mobile browsers and desktops with a retro cyber/pixel aesthetic.

Code

devanshu-puri / OutLoot

TG Trails (OutLoot) 🌿✦

Hacktoberfest Open-Source AI Challenge (Week 1: "Touch Grass")
An outdoor walking PWA where genuine physical movement unlocks virtual drops and sponsor vouchers, accompanied by private local Gemma AI missions, real-time screen time tracking, and social leaderboards.

🌟 What is TG Trails?

Most apps optimize to maximize your screen time. TG Trails is built to minimize it.

Walk outside: Put your phone in your pocket. The app uses server-authoritative physics checks to qualify genuine pedestrian movement (not GPS jumps or vehicle speeds).
Audio & Haptic Discovery: When your accumulated movement unlocks a nearby drop, your phone rings with a dual-tone chime and vibration cue.
Collect & Earn: Step within proximity of the drop to collect items across 5 rarities, backed by an append-only points ledger.
Local AI Missions & Lore: A locally-hosted open-weight model (Google Gemma via Ollama) generates tailored outdoor walking missions and immersive…

View on GitHub

GitHub Repository: https://github.com/devanshu-puri/OutLoot
License: MIT Open Source
Test Suite: 61 automated unit and integration tests passing (python -m pytest) covering physics filters, ledger accounting, authentication, trail rendering, and local Gemma fallbacks.

How I Built It

OutLoot is built with a focus on simplicity, offline resilience, and zero cloud API dependencies:

1. Architecture Overview

2. Technologies Used:

AI Engine: Google's Gemma open-weight model (gemma2:2b / gemma:7b) running locally via Ollama. Responses are strictly constrained to JSON schemas with prompt sanitization, an 8-second circuit breaker, and an offline fallback pack of 12 hand-crafted missions ensuring 100% uptime even on remote trails without internet.
Server & Logic: Python 3.12, FastAPI, and Pydantic v2 schemas. SQLite in WAL mode with strict foreign key constraints and ACID transaction managers.
Mapping: Leaflet 1.9.4 and OpenStreetMap standard tiles enhanced with CSS inversion filters (filter: invert(100%) hue-rotate(180deg) brightness(82%) contrast(108%)), delivering a clean dark cyber aesthetic without commercial API keys or tile watermarks.

* Screen-Light UX: Screen Wake Lock API and audio/haptic chimes allow players to keep the app active in their pocket at minimum screen brightness.

Why Does Open Innovation Matter?

Absolute Location Privacy: Geolocation breadcrumbs and daily movement patterns are deeply personal. Sending real-time coordinates to proprietary cloud AI providers risks leaking sensitive physical routines. By running open-weight Gemma models locally on the device or self-hosted server, your physical whereabouts never leave your custody.
Separation of Determinism and Creativity: Core game economics (points, anti-spoofing verification, vouchers) must remain 100% deterministic and auditable. Open-source architecture ensures open-weight AI is dedicated solely to creative tasks (quests, lore, and wellbeing coaching), while deterministic code guarantees fairness.
No Hidden Costs, Rate Limits, or Paywalls: Proprietary APIs inevitably deprecate endpoints, introduce per-token costs, or enforce usage caps. Open-source AI and community OpenStreetMap data make OutLoot completely free to self-host, fork, and operate forever.

4. Resilience in Nature: True outdoor exploration often happens in dead zones with zero cellular connectivity. Open-weight models and offline fallback systems ensure the game keeps working whether you are in a downtown park or a secluded mountain ridge.

My Agent Session

OutLoot was designed, implemented, tested, and deployed end-to-end with the assistance of Antigravity (Google DeepMind's agentic AI coding assistant). The session coordinated:

Drafting Pydantic schemas, database migrations, and geodesic haversine algorithms.
Developing the 61-test verification suite (pytest).
Integrating the local Gemma Ollama interface and fallback pipelines.
Automating the deployment to Render via infrastructure-as-code (render.yaml).
Interactive trail inspection and social follow mechanics. You can view the full source code and commit history directly on GitHub: 👉 https://github.com/devanshu-puri/OutLoot --- ## Prize Categories
Open-Source AI Challenge Week 1: Touch Grass (Primary Track)
Best Use of Gemma (Local open-weight Gemma model generating structured JSON quests, seasonal walking tips, and biome relic lore with offline fallback guarantees)
Best Use of Render (Automated infrastructure-as-code deployment on Render serving FastAPI backend and PWA frontend with zero external dependencies)

2. Technologies Used:

AI Engine: Google's Gemma open-weight model (gemma2:2b / gemma:7b) running locally via Ollama. Responses are strictly constrained to JSON schemas with prompt sanitization, an 8-second circuit breaker, and an offline fallback pack of 12 hand-crafted missions ensuring 100% uptime even on remote trails without internet.
Server & Logic: Python 3.12, FastAPI, and Pydantic v2 schemas. SQLite in WAL mode with strict foreign key constraints and ACID transaction managers.
Mapping: Leaflet 1.9.4 and OpenStreetMap standard tiles enhanced with CSS inversion filters (filter: invert(100%) hue-rotate(180deg) brightness(82%) contrast(108%)), delivering a clean dark cyber aesthetic without commercial API keys or tile watermarks.

* Screen-Light UX: Screen Wake Lock API and audio/haptic chimes allow players to keep the app active in their pocket at minimum screen brightness.

Why Does Open Innovation Matter?

Absolute Location Privacy: Geolocation breadcrumbs and daily movement patterns are deeply personal. Sending real-time coordinates to proprietary cloud AI providers risks leaking sensitive physical routines. By running open-weight Gemma models locally on the device or self-hosted server, your physical whereabouts never leave your custody.
Separation of Determinism and Creativity: Core game economics (points, anti-spoofing verification, vouchers) must remain 100% deterministic and auditable. Open-source architecture ensures open-weight AI is dedicated solely to creative tasks (quests, lore, and wellbeing coaching), while deterministic code guarantees fairness.
No Hidden Costs, Rate Limits, or Paywalls: Proprietary APIs inevitably deprecate endpoints, introduce per-token costs, or enforce usage caps. Open-source AI and community OpenStreetMap data make OutLoot completely free to self-host, fork, and operate forever.

4. Resilience in Nature: True outdoor exploration often happens in dead zones with zero cellular connectivity. Open-weight models and offline fallback systems ensure the game keeps working whether you are in a downtown park or a secluded mountain ridge.

My Agent Session

OutLoot was designed, implemented, tested, and deployed end-to-end with the assistance of Antigravity (Google DeepMind's agentic AI coding assistant). The session coordinated:

Drafting Pydantic schemas, database migrations, and geodesic haversine algorithms.
Developing the 61-test verification suite (pytest).
Integrating the local Gemma Ollama interface and fallback pipelines.
Automating the deployment to Render via infrastructure-as-code (render.yaml).
Interactive trail inspection and social follow mechanics. You can view the full source code and commit history directly on GitHub: 👉 https://github.com/devanshu-puri/OutLoot --- ## Prize Categories
Open-Source AI Challenge Week 1: Touch Grass (Primary Track)
Best Use of Gemma (Local open-weight Gemma model generating structured JSON quests, seasonal walking tips, and biome relic lore with offline fallback guarantees)
Best Use of Render (Automated infrastructure-as-code deployment on Render serving FastAPI backend and PWA frontend with zero external dependencies)

I Built CircleCue for a Friend, Helping them to Stay in Sync, Without Interrupting

DEVANSHU PURI — Sun, 04 Oct 2026 17:45:25 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

CircleCue is a private app that helps trusted people share the context that often leads to repeated calls and unanswered messages: what they’re doing, when they expect to be free, whether they’re travelling, and whether their phone is on silent or low on battery.

I built it for a friend,family-member and loved one who worries when I can’t answer, especially when I’m busy or travelling. I can share an update once, and they see only the details I’ve allowed them to see. CircleCue has no public feed, hidden monitoring, or background location tracking.

Demo

Live app — sign-in required
Source code on GitHub

Code

I used four files to help coding agents pick up the project across sessions:

AGENTS.md — project rules and safety boundaries
Project memory — product contracts and module status
Decision log — important trade-offs
Build log — implementation history

How I Built It

CircleCue uses Next.js and TypeScript for the web app, FastAPI and Pydantic for the API, MongoDB Atlas for storage, and Temporal for durable travel and timeline workflows. Render hosts the app and API, and Gemma runs on a DigitalOcean GPU.

Gemma turns plain-language updates like “studying till 8, no calls” into structured drafts. The user reviews a draft before it is saved or shared. The model does not decide permissions, calculate dates, or write to the database; deterministic code handles those. If the model is unavailable, a rules-based parser and regular forms keep the app usable.

I worked across coding-agent sessions, including one with Antigravity. I had limited AI tokens and no pro model, so sessions could end before the work was finished. To avoid losing decisions or asking the next agent to guess, I broke the project into modules with clear completion steps and kept the current state, decisions, and build history in the repo. I later made the handoff pack agent-neutral.

Why Does Open Innovation Matter?

An open-weight model behind a replaceable adapter gives me the option to run inference on infrastructure I control or locally during development, and to change models without rebuilding the app around a closed API.

That flexibility matters for a privacy-focused app. The model gets only the owner’s text and limited parsing context; it never sees another person’s activity. Permissions and sharing stay in the application’s deterministic code.

Open tools also helped me keep building through limited model access: I could switch agents, preserve project knowledge in versioned files, and use a rules-based fallback when model access ran out.

My Agent Session

I used Antigravity for one of the build sessions, but I don’t have that transcript saved in DevRelay yet. I’ll add a session link if I’m able to retrieve it.

Prize Categories

Best Use of Render
Best Use of DigitalOcean
Best Use of Gemma
Best Use of MongoDB Atlas
Best Use of Temporal

Friend's feedback:

CircleCue would be useful for keeping up with your class schedule and exams without having to ask you repeatedly. I could check when you’re in class or studying and see when an exam is coming up. If you’re busy, your live update lets me know you’re studying, so I can wait and connect with you later instead of interrupting you. That would be especially helpful for non-urgent things.

Designed and built by me :)

MLH hacktober fest 2026

DEVANSHU PURI — Thu, 01 Oct 2026 22:05:08 +0000

This is my first vblog , and i have came to know about this , is through mlh-hacktober-fest 2026.

DEVANSHU PURI — Thu, 01 Oct 2026 22:03:37 +0000