-
-
Smart display: five starter tiles. The Alexa+ screen is simulated; the server behind it is real.
-
Daily briefing card, health items first. Judge view puts the English under each Telugu line. Demo data.
-
Sell-day plan card with the answer in Telugu and English. The price is simulated, and the farm is fictional.
-
Health check result: urgency set by code, guidance for the vet, no diagnosis or dose; offers a vet summary. Demo data.
-
Record the day: the numbers are read back first. Nothing is saved until the farmer says yes.
-
Smart speaker mode: voice only, no screen. The answer is also written in the transcript.
-
Developer panel: the real MCP tools/call request and response behind one answer, from the public demo.
-
Architecture: Alexa+ screen and cards are simulated or sandboxed; the MCP server and its 14 tools are real. Nova is off in the demo.
-
README test line (3,486 tests) and the CI run summary of the commit pushed just before: four jobs passed, macOS manual and skipped.
-
Pilot build on a phone: Telugu home buttons and ten quick questions. Pilot link is private (real farm data); sent to judges on request.
-
Pilot build on a phone: red doctor alert when a shed's deaths pass the farmer's own limit, with a Telugu briefing. Shown with permission.
Inspiration
A broiler farmer starts the day in the shed with busy hands. Birds died overnight, and the count, the feed left and the age of each shed sit in a notebook, in memory or in a phone gallery. With several sheds, a sudden rise in deaths in one of them is easy to miss until it is costly.
CluckTalk keeps those records in one place. It tells the farmer, in Telugu and English, what needs attention first, and it answers from the farmer's own records, not from a generic guess.
A broiler farmer gave feedback during the build: they want to ask by speaking and hear the answer in Telugu. Today the answers are written Telugu. The Simulator's microphone takes English, and spoken Telugu is not built.
What it does
A farmer can do five things.
- Ask how each shed is doing. A card shows the status, with health items first.
- Get a daily briefing: what needs attention today, in order.
- Check sick birds. The farmer says what they see in one sentence, for example coughing and not eating. CluckTalk starts a check, shows "Still checking", then gives the result. Code sets the urgency. It gives guidance for the vet, never a diagnosis or a dose. The vet summary is read back first, and the SMS is simulated. It asks no symptom questions before the result; the result then offers to send a summary to the vet (Yes or No). A follow-up about the shed comes with the briefing days later (
report_health_follow_up, tested intests/contract/test_follow_up_tools.py; not shown in a short demo). - Record the day: deaths, feed, weights, treatments. The record is read back and saved only after the farmer says yes.
- Plan a sale or a feed order. The price is simulated.
The Alexa+ screen is simulated. My web app plays the Alexa+ client and speaks MCP to the real server. Its banner says: "Simulated Alexa+ screen. The backend is a real MCP server (Streamable HTTP)." The badge says "Demo data · mock host". The demo farm is fictional and has 3 houses.
How I built it
I built it alone with Claude Code. The server is a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) written in Python with the official mcp SDK. It has 14 tools. Each tool returns English text, Telugu text and a structured result, and the four cards run in a sandboxed iframe. An Agent Skill (skills/clucktalk/) describes the tools and the safety rules. The data is kept in SQLite.
A tool answers from stored state and does not wait for a model. The safety rules are in code. The urgency is set by code, so a model could raise it and never lower it.
Tests: 3,486 automated tests (524 harness, 1,757 unit, 1,205 contract). CI runs them on Ubuntu and Windows with Python 3.12 and 3.13. The macOS job is manual.
Amazon Nova 2 Lite through the Bedrock Converse API is integrated in src/clucktalk/agents/bedrock.py and is off in the live demo and in the pilot. The demo shows "Demo rules (no AI model)". Four real calls ran on 2026-09-29, outside the server, in a separate AWS account. A run through the server on 17 October failed: it was throttled, because every Nova 2 Lite quota on that account is 0 (FL-031).
Challenges I ran into
- Cards did not show on the public demo. The card sandbox origin was built as
https://localhost:with an empty port. The cards now load from a second origin under the demo's own host name, and the server checks that setting at start (FL-034, FL-037). - A check that spoken numbers equal the numbers on the card failed on the briefing date, which is display-only text. I marked it as skipped, as for the other cards (FL-035).
- One CI job failed on Windows with Python 3.12 and passed on a re-run. It is an overlapping-request test, and I have not investigated it. It is still open (FL-036).
- I could not get Amazon's Local Inspector, so I wrote my own replay of the documented flow and said so (FL-024).
Accomplishments that I'm proud of
Measured, each from a file in this repository:
- 14 tools and 3,486 automated tests (524 harness, 1,757 unit, 1,205 contract); the CI badge in
README.mdshows the latest run. - 88 of 88 golden evaluation cases pass in mock mode (run of 2026-10-16).
- Server time for the slowest step is p95 about 25 ms, on localhost in mock mode (docs/RESULTS.md). This is not Amazon's 500 ms round trip.
- In the pilot, one farmer confirmed two daily records on 19 October.
- The friction log has 34 live entries (FL-001 to FL-037).
What I learned
Put the safety rules in code and let a model only raise an alert. Label every simulated number. Read-back and confirmation suit voice well. Check Bedrock quotas with a real call, not the model list.
What's next for CluckTalk
Dairy, sheep and fish farming are planned, each needing its own vet-reviewed content. They are not started. Other next steps: Telugu speech in and out (the first request from the pilot farmer), a longer farmer pilot, a Telugu review by someone besides me, and a live AI run through the server.
How this meets each criterion
Technical Implementation. The server follows MCP 2025-11-25 over Streamable HTTP with 14 tools, and 29 protocol tests pass against it. The 3,486 tests and the 88 evaluation cases run without AWS. Limit: Nova is integrated but did not run through the server. Evidence: docs/RESULTS.md sections 1, 2 and 4; python tasks.py verify.
Design. The first screen has five starter tiles. Each answer comes in Telugu with the English under it ("Judge view"). Nothing is saved before a read-back and a yes, and every simulated number is labelled. Limit: no Telugu speech, and the layout is shown by screenshots only. Evidence: docs/evidence/2026-10-18_judge_view/; tests/contract/test_simulator_starters.py.
Potential Impact. In the pilot, one farmer used CluckTalk with two sheds and confirmed two daily records on 19 October. India's poultry market is estimated at about 2.6 lakh crore rupees (INR 2,636 billion in 2025, IMARC Group); this is a third-party market estimate, not my measurement. The target is a broiler farmer who tracks several sheds by memory or notebook. The next pilot would measure the time spent on daily records and whether deaths are noticed earlier; that is a plan, not a result. Yield, mortality, time saved and adoption are not measured. Evidence: README.md (Impact); docs/JUDGES.md.
Quality of the Idea. The idea is a helper a farmer can ask in Telugu that answers from the farmer's own records and does not diagnose. Code sets the urgency, and the reply is guidance for the vet. The schedule, safety and language layers are not poultry-specific, but other species are only planned. Evidence: docs/SAFETY_MATRIX.md; tests/unit/test_triage_rules.py.
Where to verify each claim
| Claim | Where to check it |
|---|---|
| Demo video (2:03) | https://youtu.be/wB2TxMqiE_U |
| Demo video as a file and transcript | docs/video/clucktalk_demo.mp4; docs/VIDEO_TRANSCRIPT.md |
| Farmer pilot | Link private (real farm records). Screenshots: gallery images 10-11. Available to judges on request via a Devpost message. |
| 14 tools, MCP 2025-11-25 | docs/RESULTS.md sections 1 and 4; python tasks.py verify |
| 3,486 tests (524, 1,757, 1,205) | docs/RESULTS.md section 1; the line at the top of README.md |
| CI is green | The CI badge at the top of README.md; the GitHub Actions page |
| 88 of 88 golden cases | docs/EVALS.md; docs/RESULTS.md section 2 |
| p95 about 25 ms, localhost, mock mode | docs/RESULTS.md section 5 |
| Nova integrated, four calls on 2026-09-29, off in demo and pilot | src/clucktalk/agents/bedrock.py; docs/AWS_SETUP.md sections 5.2 and 5.3; FRICTION_LOG.md FL-031 |
| India's poultry market, about 2.6 lakh crore rupees | IMARC Group, "India Poultry Market", https://www.imarcgroup.com/indian-poultry-market; docs/VIDEO_CLAIMS.md row 24. Third-party estimate, not my measurement. |
| Pilot: two sheds, two records, 19 October | docs/RESULTS.md section 9; submission.md |
| No diagnosis or dose | docs/SAFETY_MATRIX.md; tests/unit/test_triage_rules.py |
| Friction entries FL-034 to FL-037 | docs/DEV_NOTES.md; FRICTION_LOG.md |
Impact metrics
Measured: 3,486 automated tests; server time p95 about 25 ms for the slowest step, on localhost in mock mode; two daily records confirmed by one farmer in the pilot on 19 October.
Not yet measured: yield, mortality, time saved, adoption. One day of pilot data cannot show any of them.
Scale, honestly: records are kept per house (shed), not per bird. The demo farm has 3 houses and the pilot has 2 sheds. Nothing has been tested on a large farm.
Pre-existing projects
None found. The first commit (d1d1662, 2026-09-27) holds only build scaffolding (task runner, secret and cost checks, version pins, license); the product code follows in later commits. The GitHub repository was created on 2026-09-30. It was built with Claude Code.
Disclosure
Video narration: AI-generated voice (Gemini TTS).
Built with
Python 3.13, mcp 2.2.0, Model Context Protocol 2025-11-25 over Streamable HTTP, MCP Apps cards (vanilla HTML and JavaScript), Starlette 1.7.0, uvicorn 0.54.0, Pydantic 2.13.5, SQLite, Amazon Bedrock (Amazon Nova 2 Lite, Converse API), boto3 1.43.103, AWS Lightsail, AWS CLI 2.37.4, Caddy, Open-Meteo (optional, off by default), Noto Sans Telugu 2.005, Hypothesis, pytest, ruff, GitHub Actions, Claude Code.
Log in or sign up for Devpost to join the conversation.