-
-
Cover: KrishiSmriti, a voice farm advisor on Alexa+ that remembers every plot.
-
Problem: Farm history lives in people's heads, so failed treatments get repeated every season.
-
Who it's for: Built for small farmers who'd rather talk than type.
-
How farmers use it: Log a treatment, report the result, ask for advice. All by voice.
-
Memory example: In August, KrishiSmriti remembers that the June spray failed on Plot 2.
-
Generic vs KrishiSmriti: A generic advisor repeats the failed spray. KrishiSmriti remembers what happened on the plot.
-
What it remembers: Crop, problems, treatments, outcomes, timeline and follow-ups, stored for each plot.
-
Architecture: Farmer to Alexa+ to our self-hosted MCP server, with Hindsight memory and a Groq LLM.
-
Request flow: Plot history is recalled before the model answers, so advice is grounded in that field.
-
MCP tools: The tools our MCP server exposes to Alexa+.
-
Languages: English and Hindi now, with Telugu, Tamil and Marathi planned.
-
Voice design: Short answers, plain words, and a clear next step in every reply.
-
Impact: Fewer wasted sprays and a field record that builds every season.
-
Roadmap: Farmer pilot, more languages, weather and soil data, more channels.
-
Team: Shruthi Mogilicharla
Inspiration
Every season, farmers in India make the same costly guesses. A pesticide that failed on a plot last year gets sprayed again because nobody wrote it down. Advice from agri apps is generic: it knows the crop, but not your field. Most of these apps are also text-heavy, which shuts out farmers who are more comfortable speaking than reading.
We wanted an advisor that works like a good local agronomist: you talk to it, and it remembers your plot's history.
What it does
KrishiSmriti gives every plot its own memory. A farmer speaks to Alexa+ and asks something like "My cotton leaves are curling again, what should I do?" The answer takes into account what was tried on that plot before, what worked and what didn't.
- Voice first: no typing or reading needed
- Plot-level memory: treatments, outcomes, crop history and past issues are stored per plot
- Advice that learns: if a treatment failed last season, KrishiSmriti won't recommend it blindly again
- Local languages: [Hindi and English at launch; Telugu, Tamil and Marathi planned]
How we built it
- Alexa+ integration: a self-hosted MCP server exposes KrishiSmriti's tools (log a treatment, get plot history, get advice) to Alexa+
- Backend: Python and FastAPI
- Reasoning: Groq-hosted LLM for fast responses
- Memory: Hindsight stores and recalls each plot's history so advice stays grounded in what actually happened on that field
- [Hosting: e.g. AWS / Render / other]
Challenges we ran into
- [Getting Alexa+ to call our MCP tools reliably, e.g. tool naming, auth, latency]
- Keeping voice responses short enough to be useful but specific enough to be trusted
- Structuring memory so "what failed" is recalled as strongly as "what worked"
- [Language support limits]
What we learned
- [Something specific about MCP / Alexa+ agent design]
- Memory changes the conversation: once the advisor knows the plot's history, the farmer's questions get shorter and the answers get better
- [One more real lesson]
What's next
- Pilot with farmers in [district/state]
- Full Telugu, Tamil and Marathi voice support
- Linking weather and soil data to each plot's memory
- [Fire TV or other channels, if relevant]
Team
Shruthi Mogilicharla
Built With
- alexaplus
- amazon-web-services
- fastapi
- groq
- hindsight
- llm
- mcp
- python
Log in or sign up for Devpost to join the conversation.