Anthropic’s free API credits gave teams across Project Tech4Dev room to experiment with Claude and turn the best ideas into real tools for NGOs. Project Tech4Dev builds open-source platforms for NGOs, including Kaapi (AI platform), Glific (WhatsApp chatbots), Dalgo (data platform).
With these credits, we made student assessment more consistent, kept chatbots safe, helped 10–12 NGOs improve their bot prompts, let NGOs chat with their data, cut Avni app setup from weeks to days, and built AI agents that connect nonprofits with funders. Here’s how each team did it.
A) Kaapi: AI Assessment Pipeline
Many NGOs we work with run programs where students solve real problems in their communities and pitch their ideas. Judging these fairly at scale is hard. Rubrics can be vague, ideas get copied from earlier years, and a lot of the real work sits in an image or video, not a written answer.
Our pipeline first removes incomplete submissions. It then scores the rest using the NGO’s rubric and writes feedback for each student. We’re improving it together with NGO partners such as Inqui-lab Foundation, The Apprentice Project (TAP), Problem Solvers Lab (PSL), the Deodar Prize and EarthON foundation. Here’s where Claude came in.
1. Rubric refinement agent
When AI and human scores don’t match, the rubric is often the problem, not the model. A vague rubric gives different scores across human graders, between humans and AI, and even across two runs of the same model. Fixing it used to be a slow, manual loop.
So we built a Claude agent that runs the loop for us. A parser reads the rubric, a judge finds gaps and vague criteria, and a generator writes a clearer version. It stops when it hits a target (like MAE or ±1 match rate) or stops improving. Results are mixed so far, but for a few NGOs, score alignment improved by 50%, and we’re now validating the changes with them. The scoring agent itself is what runs in production today.
2. Plagiarism module
This checks if a submission is copied from another student or a known source. We used Claude to build the golden dataset of copied and original answers. The module shortlists likely matches with embeddings, then an LLM judges each one. Next, we’ll test Claude as that judge and compare results.
3. Model evals for NGOs
For TAP, PSL, Deodar Prize and Peepul, we compared Claude with other models on tough tasks, like handwritten answers with rough work and writers’ stories. Claude was among the most accurate overall, but also costlier. This tells us where the extra accuracy is worth paying for.
After these evals, TAP now plans to use Claude models for text and image data in its production pipeline.
4. Central Square Foundation (CSF) hackathon screening
For a CSF hackathon, Claude helped filter out the first 10–20% of ideas in the elimination round. Humans reviewed in parallel, so final calls stayed with people. It showed AI assessment works for screening ideas and proposals too.
B) Kaapi: Guardrails
Guardrails check every message for PII, toxicity and off-topic content, so they need to be fast and cheap. We compared fast “System 1” models with reasoning models, and the latest Claude Haiku came out accurate, consistent, fast and cheap. Full results.
C) Kaapi: Internal processes
- A set of Claude skills that help us write PRDs and SRDs.
- A PR review command that speeds up code reviews.
- A “PR ready” skill that structures PR descriptions and explains the flow of changes.
D) Glific: Prompt recommender in the Evals flow
Kaapi, the open-source AI layer behind Glific, uses Claude to help NGOs test and improve their AI-powered WhatsApp chatbots.
Most NGOs on Glific don’t have a prompt engineer. In Kaapi’s Evals flow, they test their WhatsApp bot on real user questions to see where answers fall short. Claude reviews the results and rewrites the bot’s prompt using prompt-writing guidelines. NGOs then rerun the same questions to check if answers got better. About 10–12 NGOs, including Milaan, Mowo and Sneha, used it in the last month.
E) Dalgo: Chat with your data (now live)
Since October, NGOs using Dalgo have been able to ask questions about their data in plain language, without writing SQL or relying on a data analyst. Powered by Claude Sonnet APIs, our new Chat with Data feature lets users ask questions like “How many students enrolled in Karnataka?”, get answers, and turn them into charts or dashboards within the same conversation.
More on this here
What we’ve built and are experimenting with:
- Natural language to SQL: Understands questions, identifies relevant tables and columns, generates SQL queries, and returns answers from an organization’s data.
- Conversational data exploration: Supports follow-up questions, allowing users to explore trends, compare results, and dig deeper without starting over.
- Charts, dashboards, and reports: Users can create visualizations, metrics, KPIs, and dashboards directly through chat.
- Multi-agent AI system: Built specialized agents for data analysis and platform actions, allowing the assistant to both answer questions and help users perform tasks within Dalgo.
- Human-in-the-loop safeguards: Users can review and approve queries before execution, with checks for SQL safety and sensitive data.
- AI-powered platform assistance: Users can ask how to use Dalgo features and get guidance within the same conversation.
- Accuracy and evaluation: Experimenting with automated validation, SQL correctness checks, and evaluation datasets to improve the reliability of responses.
- Expanding capabilities: Testing more complex analytical questions, improving the assistant’s understanding of organizational data, and exploring additional actions users can complete through chat.
We’re steadily refining the experience with early users and plan to expand access to all Dalgo customers next quarter.
F) Fundraising project
We’re using Claude credits to build and experiment with AI-powered tools for the Fundraising project, a platform that helps nonprofits connect with potential funders. Over the past six weeks, we’ve built AI agents and have been exploring different models, architectures, and approaches to make them more reliable, efficient, and cost-effective.
- Building AI agents: We’ve developed four agents: Ask, which answers questions about nonprofits and funders using our database; Scout, which researches organizations using publicly available information; Pitch, which helps nonprofits draft grant proposals; and a News Verifier, which checks whether news articles are relevant to specific organizations.
- Experimenting with Claude models: We’re using different model tiers depending on the complexity of the task. More capable models handle planning, reasoning, and final responses; mid-tier models handle research and reading documents; and smaller, faster models handle structured extraction, classification, and verification. We’re also experimenting with model fallbacks to balance quality, speed, and cost.
- Exploring agent architectures: We’re working with tool-calling agents, supervisor-agent patterns, parallel research agents, and multi-step workflows. We’re also experimenting with persistent state, human-in-the-loop interactions, and guardrails to make agents more reliable.
- Optimizing token usage and costs: A significant part of our experimentation has focused on prompt caching, context management, and reducing unnecessary model and tool calls. In one example, we reduced the cost of an agent run from $0.87 to $0.155 and execution time from 141 seconds to 52 seconds by improving how context and tool responses were handled.
- Building evaluations and monitoring: We’re setting up tracing, evaluation datasets, and experiments to compare agent behavior across changes. We’re tracking response quality, tool usage, token consumption, costs, and latency to understand what works and identify regressions.
- Learning where AI adds value: One important learning has been understanding when to use AI and when traditional code is more effective. For example, we moved NGO-funder matching from an LLM-based approach to deterministic scoring, making it more consistent and explainable, while continuing to use AI for research, unstructured information extraction, and natural-language interactions.
G) Avni: NGOs set up their own field apps using agentic AI
Every NGO on Avni needs the app set up for its own program: who it serves, what its field workers record, and when they should visit. Until now, an Avni specialist spent two to three weeks building each app by hand. This made onboarding slow and costly, and kept Avni out of reach for many smaller NGOs.
With our Claude API credits, we built an AI tool that reads an NGO’s program plan from a simple spreadsheet and turns it into a working app. It writes the forms and the rules behind them, and a person reviews and approves every change. NGOs in our Launchpad program now set up their own apps in two to three days, for just a few dollars in AI usage.
The credits really helped us experiment and fail faster. API spend is a variable cost and it is usually difficult to get a finance team to approve a variable budget head for experimentation. Almost everything above came out of experiments that didn’t work first but came about post iteration, experimentation and finally reaching the right design.
