What if an AI assistant could tell you exactly which law supports its answer?
That's the idea behind Seera, an AI-powered legal research platform I'm building to make Ethiopian legal information easier to explore.
Instead of relying only on an LLM's general knowledge, I built a retrieval-based approach that searches indexed legal documents and grounds answers in relevant sources.
The stack includes:
Go for the backend and streaming responses
React + TypeScript for the frontend
PostgreSQL + pgvector for legal document retrieval
RAG to connect AI-generated answers with source material
One design principle matters to me: when the available sources don't support an answer, the system should make that clear rather than pretend to know.
Seera is still a work in progress, but building it has taught me a lot about combining backend engineering, AI, and a real-world problem.
I'd test Seera with a question that retrieves a closely related law but asks for a conclusion the passage doesn't support. Does it still abstain when retrieval similarity is high?
Most AI automations fail at the inbox, not in the demo.
Problem: a founder I work with used to answer every lead by hand. Replies took hours, sometimes a day. By the time the answer arrived, the buyer had already picked a faster vendor.
Build: I shipped an agent that replies in minutes, not hours. Every reply cites its source, every action is logged, and nothing ships until it passes an eval set first.
Result: 3 AI agents now run in production for that client, with 1 paying subscriber live. The same eval-first pattern powers my other builds, including a PDF reader that returns 1 cited answer in seconds.
Offer: if replies are your bottleneck — support, sales, or ops — DM me and tell me your slowest reply. I will tell you honestly whether an agent can fix it, and what the eval would look like.