Never inventing a fact
Typed tool arguments go through all-field validation and exact-source grounding. Optional details such as ages stay unknown until they matter instead of being guessed.
سَدِيمًاArabic-first conversational AIIn development
SADEEMAN AI
A conversational assistant that asks before it assumes and cites what it finds.
An Arabic-first AI assistant for travel planning and everyday questions, with grounded web research, opt-in voice and consent-based memory on a Fastify and Next.js monorepo.
- My part
- Product design, full-stack and agent engineering
- Status
- In development
- Built with
- TypeScript
- Fastify
- Next.js
- React
- Zod
- Supabase
- PostgreSQL
- Gemini
- Groq
- Three.js
Private GitHub repositoryNo public demo — this page is the full write-up.
Actual application screenshots of the welcome screen. The background is a licensed landmark video credited inside the project.
Why it exists
Build an assistant that handles real travel planning — budgets, group size, changing constraints — without inventing facts, and answers off-topic questions plainly instead of forcing everything into a sales flow.
My part: product design, full-stack and agent engineering.
How it is built
A pnpm monorepo: apps/api is Fastify 5 with strict TypeScript and server-only provider credentials; apps/web is Next.js 16 with React 19 and Three.js; packages/contracts holds Zod 4 schemas shared by both. Supabase migrations define the schema and RLS; embedded PGlite runs the same SQL locally. The browser never holds provider or service keys — Next proxies a configured loopback API origin.
What is implemented
- One agent that chooses to ASK, ACT or ANSWER, with action-specific requirements and a single validated correction
- Gemini-primary / Groq-fallback routing with immediate failover on timeouts or quota
- Tavily research with an exact-source citation check before an answer is committed
- Multi-turn trip context: unknown optional facts stay absent and rejected proposals never become facts
- Opt-in voice: browser recording, Groq Whisper transcription, editable transcript and Gemini speech playback
- Consent-based memory; Supabase row-level security with the same SQL running locally on PGlite
- Shared Zod contracts between API and web app
- A Three.js presence driven by recording and playback, with a static fallback and reduced-motion support
Hard parts and decisions
Bounded voice input
Ten-second, independently decodable recording segments, a two-item queue and hard size limits. A provider failure stops the turn and marks partial text for review — no silent retries.
Scope discipline
When the product scope changed, the CRM routes, tools, UI and migration were removed rather than left half-alive; the retired migration was never applied to the cloud database.
Where it stands today
In developmentActive development in a private repository. The production API is deliberately kept closed until security and privacy acceptance, so there is no public deployment. SADEEMAN is intentionally not a CRM: it cannot book, pay, create customer records or contact anyone.