Hamarakhet: AI Mandi for Indian Farmers

Project Overview

Hamarakhet is an integrated, AI-powered AgriTech web platform that consolidates the most useful digital tools for Indian farmers into a single, openly accessible product. Through a clean, animated Next.js interface, farmers, buyers, and traders take part in live agricultural auctions on an open AI Mandi for Bharat, upload a photo of a diseased leaf for instant vision-based disease diagnosis, chat with a conversational AI farming assistant, view hyperlocal weather forecasts tied to their exact location, read SEO-optimised AI-generated agricultural blogs with auto-created cover art, and connect through a structured farmer community feed.

The platform is live at hamarakhet.vercel.app and was built as a final-year B.Tech project at Gargi Memorial Institute of Technology by a five-member team that I led as Team Leader and Backend Developer. The case study below walks through the problem, the market analysis we ran before scoping, the technical architecture, the engineering trade-offs, the results, and the future roadmap for the product.

The Problem We Set Out to Solve

Indian agriculture sustains nearly half of the country's workforce, yet small and marginal farmers continue to face the same set of structural problems that have persisted for decades: opaque price discovery dominated by middlemen and arhatiyas in physical APMC mandis, late and expensive access to expert crop advice, generic city-level weather data instead of field-level forecasts, and a complete absence of trusted digital community spaces for knowledge sharing. Crop diseases are frequently diagnosed only after significant yield damage has already occurred, simply because a qualified agronomist is geographically or financially out of reach.

The fragmentation problem is just as severe as the access problem. A typical farmer who wants a complete digital toolkit ends up juggling three or four separate apps: one for mandi prices, another for weather, a third for crop advisory, and a WhatsApp group for community. Most of these apps are mobile-only, several are paywalled behind subscriptions, and almost none are tuned to Indian crops, Indian languages, or Indian buying behaviour. The category was missing an open, web-first, AI-first, India-first platform built specifically for the realities of low-bandwidth Indian smartphones and the practical workflow of a working farmer.

Market and Competitor Analysis

Before scoping Hamarakhet, we ran a structured competitor analysis across the most relevant Indian AgriTech platforms to understand the gaps. The Indian AgriTech market sits at roughly USD 974 million in 2025 and is projected to grow to USD 2.52 billion by 2034 at a CAGR of around 10.59 percent, with rapid momentum behind the government's AgriStack and Digital Agriculture Mission. Despite the size of the opportunity, the category remains fragmented.

Plantix is the leading mobile-first crop disease detection app with a large library of trained diseases, but it is mobile-only, ad-supported, has no live auction layer, and offers no hyperlocal weather integration in its core flow.
DeHaat is one of India's most well-funded AgriTech platforms with a strong offline last-mile network and large input catalogue, but it depends heavily on its physical onboarding centres and operates a partly closed ecosystem with limited transparent auction-style price discovery.
e-NAM , the National Agriculture Market run by the Government of India, provides official APMC-linked price discovery, but it has no AI advisory, no leaf-photo diagnosis, no community layer, and onboarding requires going through a registered APMC mandi.
AgroStar is a mobile-first agri-commerce platform focused on selling agri-inputs (seeds, crop protection, hardware) to farmers, but it sits on the buying side of the value chain, with no live mandi auctions and no Vision-AI capability.
Kisan Suvidha , the Government of India advisory app, is free and trusted but has minimal AI integration, no leaf-photo disease diagnosis, no live auctions, and a dated interface.
Krishi Network offers an active farmer community with peer-to-peer learning, but it has no transactional layer, no Vision-AI diagnosis, and no weather forecasting.
Ninjacart, CropIn, Fasal, Intello Labs, and Farmonaut all serve adjacent niches such as supply chain digitisation, satellite analytics, and post-harvest grading, but none of them combine open auctions, vision diagnosis, weather, and community for the end farmer.

The pattern was clear. Every competitor specialises in one or two pillars of the farmer's digital life. None of them bring the full picture together for free, on the web, in a way that runs on a basic smartphone. Hamarakhet was built to fill exactly that gap.

My Role and Responsibilities

I led the project as Team Leader and primary Backend Developer. My responsibilities covered the end-to-end backend architecture, the AI orchestration layer, the deployment pipeline, and the coordination of a five-member team across backend, frontend, design, content, and SEO.

Designed the complete Supabase Postgres schema covering users, listings, auctions, bids, diagnoses, chats, blogs, community posts, comments, and the keep-alive pings table, along with row-level security policies for farmer, buyer, and admin roles.
Built the Next.js API routes and server actions that handle authentication, auction lifecycle, real-time bid placement, leaf-photo upload and inference, AI chat streaming, weather fetching, and contact-form delivery.
Integrated NVIDIA NIM for the Llama 3.3 70B conversational farming assistant and the Llama 3.2 11B Vision crop diagnosis flow, including prompt engineering tuned to Indian crops, regional pest names, and ICAR/NABARD-grade advisory tone.
Designed and built the four-stage admin-triggered AI blog pipeline on Hugging Face Inference, chaining Llama 3.3 70B for SEO metadata and long-form HTML generation, FLUX.1-schnell with an SDXL fallback for cover art, and Supabase Storage for asset persistence and CDN delivery.
Set up the Vercel deployment, environment management, secret rotation, and the daily Vercel cron heartbeat that pings a dedicated Supabase table to keep the free-tier project active.
Coordinated the five-member team across frontend, UI/UX, content, and SEO tracks, owning sprint scope, code review, and final delivery within the six-month academic window.

Solution: One Integrated AgriTech Platform

Rather than build another single-feature mobile app in a crowded category, we designed Hamarakhet as a unified, web-first product where the market, advisory, weather, and community modules feed into each other. A farmer who lists produce on the auction floor can ask the AI advisor about a suspicious leaf from the same field, check the seven-day weather outlook for the same location, and post a question to the community feed, all from a single session, on a single domain, with no app install required and no subscription fee.

The platform is positioned as an AI Mandi for Bharat: open in the sense that anyone with a browser can participate, intelligent in the sense that every transactional surface is supported by an AI layer, and Bharat-first in the sense that the wording, units, crop names, and visual language are tuned to Indian farmer expectations rather than imported from Western AgriTech.

Core Features

Live Agricultural Auctions: An Open Mandi for Bharat

The auction module is the commercial heart of Hamarakhet. Farmers list crops with quantity, unit, base price, harvest date, and location. Buyers and traders see a public feed of live auctions, place bids in real time, and watch the current highest bid and full bid history update as the auction window progresses. Every listing has a transparent paper trail in the database, which is the opposite of the closed APMC and arhatiya-driven price discovery that small farmers traditionally depend on.

The flow is intentionally web-first, runs in any modern smartphone browser, and requires no app install or APMC registration. The revenue model is a flat five percent commission on every transaction, with the platform itself remaining free for end users. This positions Hamarakhet as a direct farmer-to-buyer marketplace that complements rather than competes with the government e-NAM, and is architected to consume public Agmarknet and e-NAM datasets in a future release.

Vision-AI Crop Disease Diagnosis

The Vision-AI Crop Advisor is the feature most farmers test first. A user uploads a leaf photo from the camera or gallery; the image is sent to Llama 3.2 11B Vision hosted on NVIDIA NIM; the response is parsed into a probable disease label, a severity indication, and practical, actionable treatment guidance written in plain language. Past diagnoses are stored in the user's diagnoses table, giving each farmer a personal field history they can scroll back through across growing seasons.

Vision diagnosis is structurally the highest-value AI surface on the platform because it shifts the cost of an agronomist visit from hundreds of rupees to effectively zero, and because early detection of common crop diseases such as leaf blight, powdery mildew, septoria leaf spot, and bacterial wilt can save a significant share of a season's yield. Industry benchmarks suggest AI disease detection can flag indicators five to fourteen days earlier than the human eye.

Conversational AI Farming Assistant

The Khet AI chat assistant runs on Llama 3.3 70B through NVIDIA NIM. It handles general farming questions, government scheme guidance, pest queries, irrigation planning, soil management questions, and produce-handling advice. Responses are streamed back as rich Markdown and rendered with remark-gfm so tables, ordered lists, headings, and links all render cleanly inside the chat surface. The prompt is tuned for Indian crops, Indian regions, and an ICAR/NABARD-grade advisory tone, which gives the assistant a distinctly local voice compared to general-purpose chatbots that have not been adapted for Indian agriculture.

Hyperlocal Weather Dashboard

Weather is one of the highest-stakes variables in farming, and most generic weather apps stop at city-level forecasts. Hamarakhet uses Open-Meteo, a free, no-key weather API, to deliver a hyperlocal forecast based on the user's exact latitude and longitude. The Field Intelligence dashboard surfaces current temperature, rainfall probability, humidity, and a seven-day outlook, framed around the practical questions a farmer is actually asking: is today a good day to sow, irrigate, spray, or harvest. The dashboard also supports a Bengali microcopy track, with a longer roadmap towards multilingual coverage across Hindi, Marathi, Telugu, Tamil, and other major Indian languages.

Four-Stage AI Blog and Cover Art Pipeline

To keep the platform alive, SEO-discoverable, and educational, I built an admin-triggered blog automation pipeline at the /api/blog/generate endpoint. The pipeline runs in four stages.

Stage 1, Researcher: Hugging Face's Llama 3.3 70B is called in JSON mode to plan an SEO-optimised topic brief, meta title, meta description, slug, and tag set.
Stage 2, Writer: The same model generates a 900 to 1,200-word HTML article with internal backlinks and a parallel cinematic image prompt for the cover art.
Stage 3, Artist: FLUX.1-schnell renders a 1024 by 1024 cover image as the primary generator, with SDXL serving as a stable fallback. A strict no-humans constraint is enforced to keep the imagery culturally neutral and respectful.
Stage 4, Publisher: The image is uploaded to a dedicated Supabase Storage bucket; the blog row is inserted into the blogs table; and schema.org JSON-LD structured data is attached for rich results in Google.

The pipeline is an example of practical AI content automation: a single admin click produces a fully formatted, SEO-tagged, illustrated blog post that is immediately live on the public blog feed.

Farmer Community Feed

The Tea-Stall community module gives farmers and buyers a structured space to ask questions, share photos, post polls, and discuss practices, with tags, threaded comments, and admin moderation. It is positioned as the daily-return reason that complements the transactional auctions and the on-demand advisory features, giving the platform something for the farmer to come back to even on days when there is nothing to buy, sell, or diagnose.

Contact and Support

The contact form is wired through Zoho, with submissions routed directly into the team inbox for partnerships, feedback, and bug reports. Keeping this surface lightweight was a deliberate choice; the platform's heavy lifting lives in the AI and auction modules, not in support tooling.

Technical Architecture

Hamarakhet follows a modern serverless architecture optimised for low cost, fast cold starts, and minimal operational overhead. The Next.js application handles UI rendering, server actions, and API routes from a single TypeScript codebase. Supabase provides managed Postgres, authentication, and row-level security across all entities. Hosting runs on Vercel's edge runtime, with continuous deployment from GitHub.

The architecture intentionally uses a multi-vendor AI strategy. Conversational chat and vision diagnosis run on NVIDIA NIM. Blog text and cover art run on Hugging Face Inference. Weather is delegated to Open-Meteo, and the contact form is wired through Zoho. This split protects the platform from any single vendor pricing change or rate-limit shock and keeps the steady-state operational cost effectively at zero.

The end-to-end request flow looks like this. A user request hits the Next.js UI on the edge, passes through an API route or server action, authenticates against Supabase Auth, reads or writes against Postgres with row-level security applied, and where required fans out to the appropriate AI vendor before returning a response. A daily Vercel cron job at 03:30 UTC (09:00 IST) hits a token-protected keep-alive endpoint, which issues a single upsert to a dedicated pings table. That tiny heartbeat is what keeps the free-tier Supabase project from auto-pausing after seven days of inactivity, and is the unglamorous detail that makes the entire zero-cost model viable.

Database Design

The Supabase Postgres schema is organised around nine primary entities. The users table stores profile data and is linked to Supabase Auth for password and session management. The listings table holds crop listings, and is referenced one-to-one by the auctions table, which tracks start time, end time, current highest bid, and status. The bids table is the high-write surface, holding every bid placed across the platform with foreign keys back to auctions and users. The diagnoses table stores leaf-photo URLs, predicted disease labels, advice strings, and confidence scores. The chats table stores conversational history as JSONB, the blogs table stores Markdown content and cover image URLs, and the community_posts and comments tables power the farmer community feed.

Row-level security policies isolate the three roles. Farmers can only mutate their own listings, diagnoses, and chats. Buyers can only mutate their own bids. Admins have moderation access scoped explicitly through policy. The same Postgres instance safely serves all three personas without the application code having to enforce access control manually.

AI Architecture Deep Dive

The AI architecture is split across two host platforms by workload type. Real-time, latency-sensitive surfaces run on NVIDIA NIM, which offers managed inference for large open-weight models at predictable latency. The conversational assistant uses Llama 3.3 70B for strong general reasoning, and the leaf diagnosis flow uses Llama 3.2 11B Vision for multimodal understanding of leaf imagery.

Heavier, asynchronous workloads run on Hugging Face Inference, which offers a generous free tier well-suited to scheduled content generation. The blog pipeline uses Llama 3.3 70B for text and FLUX.1-schnell or SDXL for imagery. The split also gives the platform vendor resilience: a degradation on one provider does not take down the entire AI surface area.

Prompts on both surfaces are tuned for Indian context. The chat assistant is instructed to default to Indian crop names, Indian regional units, and ICAR/NABARD-grade tone. The vision prompt is instructed to identify common Indian crop diseases first, explain the diagnosis in plain language, and surface treatment options that are actually available to a small farmer rather than expensive imported chemistry.

SEO and Content Strategy

SEO was a first-class concern from day one, because the platform's distribution model is organic discovery rather than paid acquisition. Every page in the Next.js application is server-rendered for crawler readability. Meta titles and descriptions are tuned to high-intent farmer queries. Schema.org structured data is attached to blog posts and listing pages for rich results. The automated blog pipeline produces fresh content on a regular cadence, with internal backlinks pointing back to the auction, advisor, and weather modules to consolidate topical authority around the AI Mandi for Bharat positioning.

Keyword targeting focuses on three layers: branded queries (Hamarakhet, AI Mandi), category-defining queries (AI crop disease detection, live mandi auctions India, hyperlocal weather for farmers), and long-tail informational queries (how to treat leaf blight, best time to sow paddy in West Bengal, government schemes for small farmers). The automated content engine is built to grow the long-tail surface area over time without manual effort.

How Hamarakhet Compares to Existing Players

The comparison matrix below summarises how Hamarakhet positions against the most relevant Indian AgriTech platforms. The point of the matrix is not to claim feature parity with well-funded incumbents on every dimension, but to highlight that no other platform combines all four pillars (open market, AI advisory, hyperlocal weather, structured community) into a single free, web-first product.

Versus Plantix: Hamarakhet preserves the leaf-photo diagnosis quality through Llama 3.2 11B Vision and adds live auctions, weather, an AI chat assistant, automated blogs, and a community feed on a web platform, without ads.
Versus DeHaat: Hamarakhet is open and web-first with no onboarding gatekeeping, with transparent auction-style price discovery as a first-class feature rather than a closed channel.
Versus e-NAM: Hamarakhet is mobile-first and AI-first, designed to complement e-NAM, and is architected to integrate with public Agmarknet and e-NAM datasets in a future release.
Versus AgroStar: Hamarakhet sits on the selling side of the value chain, helping farmers sell, diagnose, and learn, rather than primarily helping them buy inputs.
Versus Kisan Suvidha: Hamarakhet keeps the same farmer-first ethos but adds modern AI features (vision diagnosis, conversational assistant), live auctions, and a modern animated web experience that still loads well on low-end devices.
Versus Krishi Network: Hamarakhet treats community as one of many features, complementing it with auctions, AI advisory, weather, and automated blogs, giving users a richer reason to return daily.

The net positioning: Hamarakhet is the only product in this benchmarked set that is web-first, free for the end user, integrated across market, advisory, weather, and community, and built on a no-vendor-lock-in AI stack.

Engineering Challenges and How We Solved Them

Keeping a free-tier database always-on. Supabase's free tier auto-pauses after seven days of inactivity, which would make the platform feel unreliable. I solved this with a token-secured Vercel cron heartbeat that performs a single upsert every 24 hours into a dedicated pings table. Steady-state cost: zero.
Keeping AI usage near zero cost at student scale. Splitting workloads across NVIDIA NIM (real-time chat and vision) and Hugging Face Inference (asynchronous blogs and image generation) keeps the platform inside the free credit envelope on both providers, with a clear pay-as-you-go path if traffic scales beyond the student-project tier.
Fair, fast live bidding on low-bandwidth connections. Auction state is read-optimised in Postgres with denormalised current_highest_bid fields on the auctions row, so the most common read (the auction feed) costs a single indexed query. Bid placement runs through tight server actions with optimistic UI updates, keeping latency acceptable even on 3G.
Multi-role access control without complex application code. Row-level security policies in Supabase isolate farmer, buyer, and admin behaviour at the database layer, which removes an entire class of authorisation bugs from the application code.
Animation quality on low-end devices. Framer Motion and Lenis.js give the marketing pages a premium, animated feel without compromising mobile performance. Animation budgets are reduced on transactional screens such as the auction floor and the diagnosis flow.
Consistent design across a five-person team. A Figma design system was locked down before frontend implementation began, with Tailwind utility classes mapped to the same design tokens. This kept the auction, advisor, weather, and community surfaces visually coherent despite being built in parallel.
Cultural and linguistic sensitivity in AI imagery. The cover-art prompt enforces a no-humans constraint, which avoids the well-known failure modes of AI image generators on South Asian skin tones, clothing, and farming attire. Imagery focuses on crops, fields, mandis, and tools instead.

Business Model and Revenue Strategy

The platform is free for end users. Revenue comes from a flat five percent commission on every successful transaction completed through the auction module. This aligns the platform's incentive with the farmer's outcome: Hamarakhet only earns when the farmer earns. The structure also avoids the access barrier that subscription-based AgriTech products create for small and marginal farmers, who are typically the most price-sensitive segment in the category.

The cost structure is engineered to make this commission rate sustainable. Hosting on Vercel's Hobby tier, database on Supabase's free tier, AI on NVIDIA NIM and Hugging Face free credits, weather on Open-Meteo, and email on Zoho's free plan together put the steady-state platform cost between zero and roughly Rs. 6,700 per month, even with AI usage scaling modestly. This makes the product viable at small scale and gives it a clear path to profitability once transaction volume grows.

Results and Impact

Shipped a production deployment at hamarakhet.vercel.app covering live mandi auctions, Vision-AI crop diagnosis, conversational AI advisory, hyperlocal weather, automated AI blogs, and a moderated farmer community in a single integrated web app.
Achieved near-zero steady-state operational cost by combining free tiers across Vercel, Supabase, NVIDIA NIM, Hugging Face, Open-Meteo, and Zoho.
Built a four-stage AI content automation pipeline that generates a fully formatted, SEO-tagged blog post with a 1024 by 1024 cover image and schema.org structured data from a single admin trigger.
Delivered the full integrated platform within a six-month academic window from June to December 2025, with a five-member team across backend, frontend, UI/UX, content, and SEO.
Architected the platform so the next phase of work, including UPI payouts, multilingual voice input, and integrations with Agmarknet and e-NAM, can ship as additive modules without re-engineering the foundation.
Produced a documented, reproducible AgriTech case study with a clear technical narrative, a defensible competitive position, and a credible long-term roadmap.

Future Roadmap

Hamarakhet was designed from the first commit as a foundation, not as a final product. The next phase of work focuses on four priorities. First, UPI payouts directly into farmer bank accounts at auction close, removing the last manual step in the transaction. Second, multilingual voice input on the AI chat assistant, starting with Hindi and Bengali and expanding outward, which would unlock the platform for farmers who type slowly or not at all. Third, deep integration with public Agmarknet and e-NAM datasets, so that benchmark mandi prices appear alongside live auction bids on the same screen. Fourth, an AgriStack-aligned data layer that lets the platform participate cleanly in the government's broader Digital Agriculture Mission as that ecosystem matures.

Key Learnings

Leading Hamarakhet pushed me well beyond pure backend work. I learned how to scope an integrated product in a category where most competitors have deliberately specialised, and to defend that integration choice with a clear market analysis rather than aspiration. I learned how to keep infrastructure costs structurally low through deliberate vendor choices and a multi-vendor AI architecture, rather than through aggressive optimisation after the fact. I learned how to coordinate a multi-disciplinary five-person team where backend, frontend, design, content, and SEO needed to ship in lockstep across a six-month window. Most importantly, I learned that for an Indian farmer audience, the win is not the cleverest model or the slickest animation; it is the platform that quietly works on a basic smartphone, in a noisy mandi, on a weak 4G signal, in the user's own language, for free.

Live Project

Visit Hamarakhet at hamarakhet.vercel.app to see the platform in action, or explore my other work on my portfolio .