Offline (Hosting Paused)Role: Creator & Engineer

Source Mind

Embeddable AI assistant trained on website, document and learning materials

An embeddable AI assistant capable of delivering answers grounded in website content, documentation, or educational curricula via a lightweight script tag.

Project is not currently hosted due to hosting funding constraints. Architecture and specs documented below.

FastAPIPythonReactMySQLOpenAI APIDeepSeek APIClaude API

1. Context & The Core Problem

Organizations and educators needed a way to embed specialized conversational assistance that strictly adhered to their own authoritative materials (such as step-by-step math curricula) rather than hallucinatory open models.

The Bottleneck:General AI chatbots frequently hallucinate or deviate from specific textbook methodologies, and existing commercial solutions were expensive and hard to embed cleanly into custom web pages.

2. Engineering Solution & Implementation

Built a grounded Q&A engine using FastAPI with multi-provider LLM integrations (OpenAI, Claude, DeepSeek) and developed an embeddable React chat widget deployable via a simple JavaScript snippet.

Direct Contributions & Ownership

Developed high-performance asynchronous API endpoints using Python FastAPI for token streaming and conversation memory.
Implemented content-grounded prompt templates ensuring answers adhered strictly to supplied syllabus rules (e.g. step-based mathematics).
Engineered an embeddable, customizable React widget supporting light/dark embedding and dynamic theming.
Integrated multi-provider LLM failover between Claude, OpenAI, and DeepSeek for cost efficiency and reliability.

3. System Architecture & Boundaries

System Architecture

Sanitised conceptual schematic representing components and integration boundaries

Sanitised Schematic

Embeddable Chat Widget

Injected into client sites via script tag

FastAPI Backend

Context injection, guardrails and prompt grounding

MySQL Storage

Conversations, tenant API keys and configuration

LLM APIs (Claude / OpenAI / DeepSeek)

Multi-model completion and reasoning layer

Data Flow & Interactions

Embeddable Chat Widget→ Streaming SSE →FastAPI Backend

FastAPI Backend→ Audit & Config →MySQL Storage

FastAPI Backend→ Grounded Context →LLM APIs (Claude / OpenAI / DeepSeek)

4. Outcomes & Verified Impact

Successfully demonstrated step-by-step school math problem solving verified against provided syllabus logic.

Evidence level: user-reported

Enabled third-party site owners to integrate an AI knowledge assistant in under 5 minutes with a single script tag.

Evidence level: user-reported

5. Key Challenges & Architectural Lessons

Engineering Challenges

•Grounding answers tightly to prevent hallucinations when dealing with deterministic domain problems like mathematical formulas.
•Balancing token costs across multiple LLM providers while keeping latency low for interactive chat.

Takeaways & Reflection

•Strict grounding rules and fallback prompting are essential when AI assistants are exposed to educational users.
•Decoupling the frontend widget from specific LLM providers allows seamless switching as pricing and performance evolve.

All Case Studies Discuss this implementation