DEV Community: Pratyush Panda

SlothFridge: The Couch-Potato Culinary Copilot (TabPFN + Gemma 2 + ElevenLabs)

Pratyush Panda — Wed, 21 Oct 2026 09:08:07 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

It is 10:45 PM. You are lying completely flat on your bed, scrolling on your phone, starving.

You drag yourself to the kitchen, pull open the fridge door, and stare into the void: two cold eggs, a dying heel of stale bread, half a bowl of curd (dahi), and a stray slice of cheese.

The traditional internet recipe requires chopping three shallots, preheating an oven, using four prep bowls, and dirtying a cutting board. Opening a food delivery app means spending ₹250 and waiting 45 minutes while fighting decision fatigue. Most people just give up and go to sleep hungry.

SlothFridge is the couch-potato culinary copilot built to honor International Sloth Day:

The Couch Ingestion Check: Tap whatever random items are dying in your fridge without leaving bed or typing out complex measurements.
Tabular Fatigue & Standing Regression (TabPFN): Evaluates kitchen variables against 130 empirical lazy-cooking observations to predict your active standing minutes, classify cleanup fatigue tiers (Zero Effort (Mug), Low Effort (1 Pan), High Effort (Avoid)), and calculate your sink debt penalty.
Lazy Recipe Synthesis (Gemma 2 via Groq): gemma2-9b-it crafts an unapologetically low-effort 3-step meal designed around one non-negotiable rule: zero cutting boards, zero knives, and you eat directly from the vessel so you never look at the sink.
Sleepy Audio Walkthrough (ElevenLabs): Streams hands-free audio instructions using eleven_turbo_v2_5 in a slow, deadpan voice, guiding you through low-heat cooking so you can retreat back to bed while it cooks.

Demo

Live Web Application: slothfridge.vercel.app

Local / Self-Hosted Backend: FastAPI + Uvicorn service running TabPFN CPU inference and ElevenLabs streaming.

The interface is built with a mid-century warm stationery aesthetic (#EFE9DD ground, #141C2B carbon ink, and Newsreader + Courier Prime typography), featuring self-drawing effort curves and zero-border-radius precision.

Code

The entire open-source repository—including the reference dataset, TabPFN dual inference engine, Gemma 2 prompt templates, and multi-page editorial frontend—is public on GitHub:

Pratyush-Panda-2006 / SlothFridge

The couch-potato culinary copilot. Predicts cleanup fatigue and active standing minutes via TabPFN, crafts ultra-lazy 1-pan/mug recipes with Gemma 2, and streams deadpan hands-free cooking audio with ElevenLabs for Hacktoberfest Week 3.

SlothFridge: The Couch Potato Culinary Copilot 🦥🍳

Built for the Hacktoberfest Open-Source AI Challenge Week 3
Theme: Slow and Steady / International Sloth Day

SlothFridge takes expiring groceries entered by exhausted users lying on their beds and outputs an uncompromisingly lazy, 1-pan or microwave recipe. It uses TabPFN to predict cleanup fatigue and continuous standing minutes, Gemma 2 (via Groq) to synthesize lazy recipes, and ElevenLabs to stream slow, deadpan audio walkthroughs.

🏛 Architecture & Open Stack

TabPFN Fatigue & Standing Time Engine (services/tabpfn_service.py):
- Trained on data/sloth_meals.csv (130 realistic kitchen observations).
- Predicts active standing minutes (continuous regression) and cleanup fatigue tier (Zero Effort (Mug), Low Effort (1 Pan), High Effort (Avoid)).
- Computes utensil penalty and inverted Sloth Approval Score (0–100, where higher is lazier).
Gemma 2 Couch Potato Chef (services/gemma_service.py):
- Synthesizes lazy recipes using gemma2-9b-it on Groq with strict JSON…

View on GitHub

How I Built It

SlothFridge connects tabular foundation inference with open-weight language synthesis and streaming speech:

1. TabPFN Fatigue & Standing Time Engine (`services/tabpfn_service.py`)

Rather than relying on vague heuristics or heavyweight neural training loops, SlothFridge utilizes Prior Labs' TabPFN fitted on in-memory tabular data (data/sloth_meals.csv):

Continuous Regression: Estimates exact standing time down to the decimal (e.g., 3.8 minutes).
Multi-Class Probabilities: Computes calibrated posterior distributions across Zero Effort (Mug), Low Effort (1 Pan), and High Effort (Avoid).
Cognitive Anomaly Detection: Flags an alert if a user inputs complex ingredients requiring extensive knife work while selecting "Couch Comatose" commitment.

2. Gemma 2 Couch Potato Chef (`services/gemma_service.py`)

Using Google's open-weight gemma2-9b-it served through Groq, the engine acts as an exhausted, realistic chef. It strips out all culinary pretense:

Enforces hand-tearing over knife-chopping.
Mandates eating straight from the skillet or mug with a fork.
Generates deadpan, 3-step instructions tailored to your available heat source.

3. ElevenLabs Sleepy Voice Narration (`services/voice_service.py`)

Using eleven_turbo_v2_5 with the calm, deadpan "George" voice profile, the app synthesizes low-latency audio walkthroughs. Users start the audio, dump ingredients in, set the heat to low, and return to bed until the timer sounds—touching neither a cutting board nor their phone screen with messy hands.

Why Does Open Innovation Matter?

Zero-Toll Everyday Utilities: A consumer app for lazy home cooking cannot survive behind expensive proprietary API paywalls or ₹800/month subscription gates. Open models like Gemma 2 and CPU-efficient tabular models like TabPFN run at near-zero inference cost.
Local Sovereignty: Your kitchen habits, pantry ingredients, and late-night routines stay private.
In-Context Tabular Intelligence: TabPFN's transformer-based approach allows developers to evaluate small, curated tabular datasets without managing hyperparameter tuning pipelines or cloud database clusters.

The Lazy Payoff

To test the real-world payoff, I simulated a late-night cooking scenario using a typical hostel desk stash: 2 eggs, half an onion, two slices of stale bread, and a slice of processed cheese.

The Traditional Recipe: 22 minutes of standing, 1 chef's knife, 1 cutting board, 1 whisk, 1 prep bowl, 1 frying pan, and 1 plate. Total sink debt: 7 items.
The SlothFridge Payoff:
- TabPFN Predicted Standing Time: 4.2 minutes.
- Cleanup Fatigue Tier: Low Effort (1 Pan) (94.2% posterior confidence).
- Gemma 2 Generated Recipe: "The 1-Skillet Bread-Egg Slump" — Tear the stale bread by hand into a cold pan, crack both eggs over the bread, crush the onion half into rough chunks with your palm, drop the cheese on top, cover with a lid, and walk back to bed for 5 minutes. Eat directly from the pan with a single fork.
- Actual Utensils Washed: 1 skillet, 1 fork.
- Effort Saved: ~18 minutes of standing avoided and an 85% reduction in dirty dishes.

Prize Categories

Best Use of TabPFN
Best Use of Gemma
Best Use of ElevenLabs
Best Use of Vercel

HostelBites: Zero-Shot Tabular Spoilage Arbitrage & Midnight Flash Combos for Campus Stalls

Pratyush Panda — Fri, 16 Oct 2026 18:17:33 +0000

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

What I Built

Outside college campus gates in India (like ITER/SOA, KIIT, and OUTR in Bhubaneswar), late-night roll counters, burger stalls, and small cloud kitchens rely heavily on hungry hostel students. However, their razor-thin margins face a daily threat: unpredictable perishable inventory waste.

Freshly marinated chicken, paneer cubes, and burger buns spoil overnight under humid coastal conditions. If it suddenly rains at 9:00 PM, or if hostels enforce quiet-hour study gates before mid-semester exams, stall owners routinely throw out ₹1,000 to ₹2,500 worth of fresh stock by morning.

HostelBites is an autonomous inventory liquidation copilot built for the daily 8:00 PM kitchen check-in:

The 8:00 PM Inventory Audit: The line cook or stall owner enters remaining raw stock weights alongside environmental conditions (exam week flags, rain, day of the week).
Zero-Shot Tabular Spoilage Prediction (TabPFN): Ingests the tabular kitchen metrics to classify spoilage risk, outputs posterior class-probability distributions, and calculates continuous clearance windows and mandi financial loss exposure.
Campus-Targeted Flash Combos (Gemma 2): Uses gemma2-9b-it via Groq to synthesize high-margin flash combo meals that bundle surplus stock, generating ready-to-broadcast WhatsApp copy tailored with local campus slang and exam urgency.
Hands-Free Kitchen Audio Briefing (ElevenLabs): Converts the preparation priorities into a clear Hinglish voice walkthrough using the Turbo engine, letting the line cook hear batch-fry priorities over loud kitchen exhaust fans.

Stall owners gladly pay a modest ₹199/month micro-subscription because salvaging just one evening's worth of unsold chicken pays for the service for an entire semester.

Demo

Live Application: https://hostelbitesapp.vercel.app

Code

The entire open-source repository—including the tabular prior dataset, TabPFN foundation pipelines, Gemma 2 prompt engineering, and the Tech Editorial frontend—is public on GitHub:

Pratyush-Panda-2006 / HostelBites

Zero-shot tabular food waste arbitrage engine for campus cloud kitchens. Predicts midnight perishable loss via TabPFN, synthesizes student flash deals via Gemma 2, and streams line-cook audio walkthroughs with ElevenLabs.

HostelBites 🌙: Midnight Food Arbitrage Engine

Production-ready AI waste prediction & dynamic flash deal engine for late-night college campus food stalls, roll joints, and cloud kitchens.

⚡ Overview

Late-night food stalls outside engineering campus gates (such as ITER / SOA, KIIT Food Street, Silicon Tech, and Manipal) face a persistent operational challenge: unpredictable closing surplus. Unsold chicken, paneer, and breads face morning spoilage and financial write-off.

HostelBites solves this by bridging tabular machine learning and generative campus marketing:

The 8:00 PM Kitchen Check: Takes remaining raw inventory (chicken in kg, paneer in kg, buns/parathas) alongside environmental signals (exam week status, rain, day of the week).
TabPFN Spoilage Inference: Employs TabPFN (CPU Prior-Data Fitted Network) trained on 120 days of campus operational records (data/midnight_orders.csv) to predict risk level (High, Moderate, Safe) and compute potential financial exposure at mandi…

View on GitHub

How I Built It

HostelBites links tabular machine learning with generative copy and neural speech synthesis:

Prior Labs TabPFN-3.5 Foundation Model: Rather than training a brittle, heavy gradient-boosted tree that requires thousands of samples, TabPFN functions as an in-context tabular foundation model. Fitted directly over 150 nights of campus kitchen records (data/midnight_orders.csv), it delivers:
- Categorical Risk Classification: Classifies risk into Safe, Moderate, High, and Critical.
- Calibrated Posterior Probabilities: Provides confidence percentages across every risk tier.
- Continuous Clearance Windows: Projects remaining liquidation hours before demand drops to zero.
- Outlier Detection: Flags dangerous inventory build-ups (e.g., surplus poultry during sudden storms).
Gemma 2 (gemma2-9b-it via Groq): Analyzes the inventory surplus and risk status to generate campus-themed WhatsApp broadcasts with emojis, gate delivery instructions, and combo names ("End-Sem All-Nighter Double Roll").
ElevenLabs Neural Turbo (eleven_turbo_v2_5): Synthesizes practical 30-second Hinglish voice instructions with the George voice profile, backed by automatic Web Speech API browser fallbacks.
Frontend Architecture & Hosting: Styled with a "Tech Editorial" aesthetic (Playfair Display, Space Grotesk, and Space Mono) on a 4-column guide grid with scan-line progress indicators, deployed on Vercel.

Why Does Open Innovation Matter?

Open-weight models and open foundation models are essential for micro-enterprises:

Protecting Micro-Margins: Proprietary closed APIs charge high token fees that consume the small margins of street vendors. Open-weight models like Gemma 2 and CPU-efficient foundation models like TabPFN run at near-zero inference cost.
Data Sovereignty & Local Control: Daily inventory ledgers, raw purchase costs, and local customer logs stay private within the vendor's application rather than being harvested by commercial cloud providers.
Zero-Shot Adaptability: TabPFN's transformer architecture evaluates small tabular datasets directly in memory without hyperparameter tuning, bringing institutional-grade yield optimization to grassroots developers.

My First User

To test HostelBites in the real world, I visited a popular roll and paratha stall near the engineering student hostels in Jagamara, Bhubaneswar around 8:30 PM.

The cook had approximately 3.5 kg of pre-marinated chicken and 30 dough balls remaining on a quiet evening:

TabPFN Output: Classified the situation as Moderate risk, computing a 2.5-hour clearance window and ₹1,415 in at-risk stock.
Gemma 2 Output: Formatted an urgent "Exam Midnight Craving Alert: Jumbo Butter Chicken Roll @ ₹119" WhatsApp broadcast.
Feedback: The owner explained that refrigerating opened, marinated meat degrades its texture for the next day. He forwarded the broadcast into two local student hostel groups. Within 40 minutes, the stall received 9 orders, clearing over 2 kg of chicken before midnight. He affirmed that a ₹199/month fee is an easy investment if it reliably clears perishable stock during exam seasons or rainy days.

Prize Categories

Best Use of TabPFN
Best Use of Gemma
Best Use of ElevenLabs
Best Use of Vercel

CampusBahar: Touch Grass Before Monday Classes — Zero-Screen-Time Trail Scout

Pratyush Panda — Fri, 09 Oct 2026 04:55:16 +0000

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

What I Built

Engineering and college students across Bhubaneswar (from campuses like SOA/ITER, KIIT, OUTR, AIIMS, and Silicon) often spend entire weekends inside air-conditioned hostel rooms scrolling feeds or grinding code blocks. Many migrant hostellers want to explore, but lack of local transit clarity, intense coastal humidity windows, and Odia language barriers keep them trapped indoors.

CampusBahar (PravasiTrail) is an intentional anti-screen exploration engine. Its mission is to be used for less than 30 seconds:

Hostel-to-Trail Routing: Select your campus hub or auto-detect your location with GPS to discover 18 natural, forest, and ancient rock-cut destinations across Odisha.
TabPFN Diurnal Climate Windows: Predicts heat, humidity, and UV solar comfort using a prior-data fitted tabular model so students step out only during safe dawn or dusk hours.
4-Way Transparent Transit Matrix: Real-time fare comparisons across Mo Bus routes, shared autos, bike taxis, and split cabs.
Hands-Free Audio Walk (ElevenLabs): Converts the pocket guide into clear audio so students can put their phones away, put in their earphones, and navigate purely through sight and touch.
Colloquial Street Odia: Provides authentic Odia script, phonetic transliteration, and street tips to negotiate with local auto drivers and bus conductors without stress.

Demo

Live Application: https://campusbahar.vercel.app
Production Backend API: https://campusbahar.onrender.com
API Health Check: https://campusbahar.onrender.com/api/health

Code

The entire open-source repository (FastAPI backend + Memphis postmodern UI) is available on GitHub:

Pratyush-Panda-2006 / CampusBahar

A zero-screen-time weekend exploration engine for Bhubaneswar students. Built with TabPFN tabular models and Groq Gemma 2 to check coastal heat windows, transit fares, and trail guides before stepping outside.

How I Built It

CampusBahar is built around a lightweight, privacy-focused open-weight AI architecture:

TabPFN Foundation Model (tabpfn): We fitted TabPFN on diurnal coastal microclimate cycles (bbsr_weather_comfort.csv). Unlike heavy ML pipelines that require hyperparameter tuning or training iterations, TabPFN delivers instant zero-shot tabular inference directly on CPU, classifying hours into Optimal, Moderate, or Avoid conditions based on temperature, relative humidity, and UV index.
Gemma 2 (gemma2-9b-it via Groq): Synthesizes concise 4-point micro field cards (Transit Route, Sensory Touch-and-See Checkpoint, Driver Phrase in Odia, and Digital Detox Rules) tailored specifically for non-Odia students from Jharkhand, Bihar, and UP.
Google Gemini API (gemini-2.5-flash): Powers the bilingual Odia translation engine, handling colloquial street phrasing and phonetic transliteration to demystify transit communication.
ElevenLabs Speech API (eleven_turbo_v2_5): Generates lightweight audio walk narrations with the Rachel voice model, allowing students to disconnect from screens entirely once on the trail.
Cloud Infrastructure: The asynchronous FastAPI backend is deployed on a Render Web Service container in Singapore, and the frontend is hosted on Vercel with pure static edge rewrites.

Why Does Open Innovation Matter?

Open-source and open-weight models give developers the power to solve hyper-local civic and student challenges without being tied to expensive, closed enterprise platforms.

Using TabPFN and Gemma 2 means community tools like CampusBahar can run inference with minimal latency and predictable overhead. Open innovation allows models to be adapted to local micro-climates, regional bus networks, and campus realities, giving students an accessible, open-weight toolkit that encourages them to close their laptops and step outside into the real world.

Prize Categories

Best Use of TabPFN
Best Use of Gemma
Best Use of ElevenLabs
Best Use of Render

LabExplain: A Zero-Login Code Tutor for University Labs Powered by Gemma 2

Pratyush Panda — Fri, 02 Oct 2026 18:42:36 +0000

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

What I Built

During university computer programming labs, instructors often demonstrate syntax implementations (like Python file handling or pointers in C) and quickly move ahead. My friend and lab partner struggled to grasp the internal mechanics of the context manager (with open(...) as file:) and list comprehensions during a recent class.

He wanted to ask an AI tutor for an immediate line-by-line explanation, but doing so on shared university terminals required signing into personal Google or OpenAI accounts. Logging into personal accounts on public lab workstations presents serious privacy risks, including cached credentials, open sessions, and browser data leakage.

To solve this, I built LabExplain: a zero-login, peer code tutor designed specifically for shared lab environments. Students do not need individual accounts, OAuth logins, or credentials. Instead, any student at a shared machine types a 6-digit session PIN (123456), pastes their lab snippet, and receives an instant, line-by-line pedagogical breakdown.

Demo

Live Deployed App: https://labexplain-ai.onrender.com
Session PIN for Testing: 123456

Code

The complete source code is public and open-source:

Pratyush-Panda-2006 / LabExplain-AI

A privacy-first code explainer built with FastAPI and Google Gemma 2 open weights. Allows students on shared lab computers to get instant line-by-line breakdowns for Python, C++, and Java using a simple 6-digit session PIN.

⚡ LabExplain

Zero-login, peer-accessible university code tutor powered by Google Gemma 2 (gemma2-9b-it) via Groq Cloud.
Built for the DEV Hacktoberfest Weekend Challenge.

🎯 Problem Statement

In university coding labs, undergraduate students often encounter unfamiliar syntax and cryptic runtime errors (e.g., Python context managers, recursion base cases, pointer dereferencing in C/C++). However, students cannot safely use commercial AI tools on public lab machines because doing so requires logging into personal accounts, creating serious credential exposure and session leakage risks on untrusted workstations.

💡 The Solution

LabExplain is a zero-login, peer-accessible code tutor:

Students sit down at any university lab computer.
Enter the active 6-digit lab session PIN (482910 or set by the instructor/TA).
Paste their confusing code snippet.
Receive an encouraging, line-by-line pedagogical breakdown powered by Google's open-weight Gemma 2 (gemma2-9b-it) running at ultra-low latency on Groq Cloud.

🏗️ Architecture & Project Layout

labexplain-ai/
├──

…

View on GitHub

How I Built It

Open-Source AI Inference: Powered by open weights served via Groq's low-latency inference engine for near-instant responses during lab classes.
Backend Architecture: Built with FastAPI and Pydantic, featuring session PIN verification, language routing (Python, C, C++, Java, JavaScript, SQL), and structured pedagogical prompt engineering.
Frontend Design: A single-page dawn-lake glassmorphism UI featuring responsive CSS tokens, custom dark-glass dropdown components, and an expandable line-by-line modal.
Cloud Hosting: Deployed as a web service on Render.

Why Does Open Innovation Matter?

Open-weight models and open innovation are essential for this architecture:

Privacy on Shared Hardware: Proprietary closed-source platforms mandate individual user profiles and track activity. Open-weight models decouple intelligence from personal accounts, allowing lightweight, PIN-brokered sessions on public lab computers.
Campus LAN & Air-Gapped Readiness: Because the underlying weights are open, educational institutions can host the model directly on local campus servers or offline intranets where student machines lack external internet access.
Accessibility: It ensures all students get high-quality learning assistance without paywalls or individual subscriptions.

Prize Categories

Best Use of Gemma
Best Use of Render

DEV Community: Pratyush Panda

SlothFridge: The Couch-Potato Culinary Copilot (TabPFN + Gemma 2 + ElevenLabs)

What I Built

Demo

Code

Pratyush-Panda-2006 / SlothFridge

The couch-potato culinary copilot. Predicts cleanup fatigue and active standing minutes via TabPFN, crafts ultra-lazy 1-pan/mug recipes with Gemma 2, and streams deadpan hands-free cooking audio with ElevenLabs for Hacktoberfest Week 3.

SlothFridge: The Couch Potato Culinary Copilot 🦥🍳

🏛 Architecture & Open Stack

How I Built It

1. TabPFN Fatigue & Standing Time Engine (services/tabpfn_service.py)

2. Gemma 2 Couch Potato Chef (services/gemma_service.py)

3. ElevenLabs Sleepy Voice Narration (services/voice_service.py)

Why Does Open Innovation Matter?

The Lazy Payoff

Prize Categories

HostelBites: Zero-Shot Tabular Spoilage Arbitrage & Midnight Flash Combos for Campus Stalls

What I Built

Demo

Code

Pratyush-Panda-2006 / HostelBites

Zero-shot tabular food waste arbitrage engine for campus cloud kitchens. Predicts midnight perishable loss via TabPFN, synthesizes student flash deals via Gemma 2, and streams line-cook audio walkthroughs with ElevenLabs.

HostelBites 🌙: Midnight Food Arbitrage Engine

⚡ Overview

How I Built It

Why Does Open Innovation Matter?

My First User

Prize Categories

CampusBahar: Touch Grass Before Monday Classes — Zero-Screen-Time Trail Scout

What I Built

Demo

Code

Pratyush-Panda-2006 / CampusBahar

A zero-screen-time weekend exploration engine for Bhubaneswar students. Built with TabPFN tabular models and Groq Gemma 2 to check coastal heat windows, transit fares, and trail guides before stepping outside.

How I Built It

Why Does Open Innovation Matter?

Prize Categories

LabExplain: A Zero-Login Code Tutor for University Labs Powered by Gemma 2

What I Built

Demo

Code

Pratyush-Panda-2006 / LabExplain-AI

A privacy-first code explainer built with FastAPI and Google Gemma 2 open weights. Allows students on shared lab computers to get instant line-by-line breakdowns for Python, C++, and Java using a simple 6-digit session PIN.

⚡ LabExplain

🎯 Problem Statement

💡 The Solution

🏗️ Architecture & Project Layout

How I Built It

Why Does Open Innovation Matter?

Prize Categories

1. TabPFN Fatigue & Standing Time Engine (`services/tabpfn_service.py`)

2. Gemma 2 Couch Potato Chef (`services/gemma_service.py`)

3. ElevenLabs Sleepy Voice Narration (`services/voice_service.py`)