Workflow Executor Backend

A scalable, LLM-assisted workflow execution backend designed to convert unstructured user problems into structured decisions and deterministic backend actions.

This project is currently being developed as part of an internship at Jio Platforms Ltd, with a strong focus on backend correctness, safety, and scalability.
The architecture is intentionally product-oriented so it can later be extended, generalized, and presented as an independent platform.

---

To see the Architecture up untill phase 7 view /docs/ARCHITECTURE.md

To view the invariants : /docs/INVARIANTS.md

..huh

1. Core Idea

The system enforces a strict separation of responsibilities:

1. LLMs are used only for intent and entity extraction
2. Schemas (Pydantic) enforce structure and validation
3. Backend code controls all decision-making and execution
4. APIs expose only final outcomes, never internal reasoning

This approach ensures predictability, auditability, and production readiness.

---

2. High-Level Architecture

Client (Frontend) | v FastAPI Backend | v Schema-Aligned LLM Extraction | v Pydantic Validation | v Deterministic Decision Logic | v Tool / API Execution | v Structured API Response

The backend is designed first, with the frontend layered on top later.
This enables clean API contracts, easy frontend integration, and independent backend evolution.

---

3. Design Principles

• LLM output is treated as untrusted input
• All boundaries are enforced via schemas
• Decision logic never lives inside prompts
• Failures are contained at validation boundaries
• New workflows are added via schemas and tools, not prompt hacks

---

4. Scalability and Product Vision

Although currently scoped for internal workflows at Jio Platforms Ltd, the system is designed as a generic workflow execution platform.

Planned scalability includes:

• Multiple workflow types
• Domain-specific extraction schemas
• Pluggable tool and API layers
• Admin and client-facing frontend applications

The backend remains the single source of truth, with all future frontends consuming it via stable API contracts.

---

5. Tech Stack

• Python 3.x
• FastAPI
• Pydantic
• LangChain

---

6. Project Status

Current focus:

• Backend-first architecture
• Safe and constrained LLM integration
• Deterministic workflow execution
• Extensibility and clarity over feature breadth

Frontend layers and advanced agent behaviors will be added in later stages.