Srot Core

Srot-core is an AI-powered engineering intelligence system that turns meetings into code-aware, execution-ready developer tasks.

It bridges the gap between what is discussed in a meeting and what an engineer (or AI agent) needs to actually implement it.

Feature	Note-Taker Apps	Srot Core
Context	Transcript only	Transcript + Codebase
Output	Summaries	Summaries + Execution-ready tasks
Accuracy	Generic / sometimes misleading	Grounded in actual code
Post-Meeting Work	Manual (create tickets, find code)	Automated (Jira + task files + CLI)

---

What It Does

Engineering teams lose significant time translating meeting discussions into actionable tickets, and those tickets rarely contain enough context for an engineer to start work immediately. srot-core solves this end to end:

1. Streaming Intelligence: Real-time transcription with context resolution (handles noise, pronouns, mixed language).
2. Context-Aware Extraction: Detects actionable items and avoids irrelevant chatter.
3. Jira-Ready Outputs: Generates structured, non-duplicate tickets with proper context.
4. Codebase Enrichment: Maps tasks to actual services, functions, entities using Neo4j + Qdrant.
5. Dev-Ready Task Files: For every task, a structured markdown file is generated for use by engineers or AI coding agents.

---

How It Works

Intelligence Pipeline

Audio Stream (PCM 16kHz)
        │
        ▼
  Deepgram ASR
  (Hindi support, interim results)
        │
        ▼
  Confidence & Rule Filter
  (drops noise, confirms quality)
        │
        ▼
  Context Manager & Resolver
  (tracks project/module, enriches text)
        │
        ▼
  LLM Trigger (Gemini)
  (decides if a task is being discussed)
        │
        ▼
  Task Extraction & Jira State Builder
  (shapes tasks with code context & domain packs)
        │
        ▼
  Hybrid Summary Generation
  (LLM-corrected cumulative markdown)
        │
        ▼
  Dev Layer
  (Structured task files for engineers/agents)

Code Knowledge

Srot-core indexes TypeScript/NestJS codebase into a hybrid knowledge store:

• Neo4j:

  • Nodes: Project, Service (Controller, Resolver, Entity, Module), GraphQLType, Interface, APIEndpoint, Function, Enum, EntityColumn, and DomainEntity.
  • Relationships: DEFINED_IN, TRIGGERS, USES, HANDLES, DEPENDS_ON, EXTENDS, IMPLEMENTS, HAS_COLUMN, REFERENCES.

• Qdrant: A multi-vector store (Semantic + Name vectors) providing a high-fidelity memory of code intent and identity. All projects share a single collection with metadata-level isolation.

At task generation time, both stores are queried: graph lookups find exact structural matches (service names, enum values, function signatures), while vector search finds semantically relevant code chunks. The combined context feeds the LLM.

Dev Task Files

Each task produces a file like dev_tasks/m1-t2.md containing:

• Project name (resolved from graph)
• Domain entity and task type
• Affected services and APIs (with relative file paths)
• Existing state (concrete code snippets/logic from context)
• Step-by-step required changes
• Constraints and unknowns
• Validation approach
• Execution hint for AI agents

---

Tech Stack

Layer	Technology
Backend	FastAPI
Frontend	Streamlit
LLM	Gemini 2.5 Flash Lite
Transcription	Deepgram Nova-2 (Streaming & Batch)
Embeddings	Gemini Embedding 001
Graph database	Neo4j
Vector database	Qdrant
Relational database	MySQL
Code parsing	Tree-sitter (TypeScript Grammar)
Audio processing	static-ffmpeg + pydub
Jira integration	Jira REST API v3

---

Prerequisites

• Python 3.12+
• MySQL
• Neo4j
• Qdrant
• ffmpeg

---

Setup

1. Create virtual environment

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Configure environment

Create a .env file in the project root:

# MySQL
MYSQL_HOST=127.0.0.1
MYSQL_PORT=3306
MYSQL_USER=your_mysql_user
MYSQL_PASSWORD=your_mysql_password
MYSQL_DATABASE=srot_core

# Neo4j
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_neo4j_password

# Qdrant
QDRANT_URL=http://localhost:6333
QDRANT_API_KEY= 

# Jira
JIRA_BASE_URL=https://your-domain.atlassian.net
JIRA_EMAIL=your_email
JIRA_API_TOKEN=your_jira_api_token
JIRA_PROJECT_KEY=your_project_key

# AI
GEMINI_API_KEY=your_gemini_api_key
DEEPGRAM_API_KEY=your_deepgram_api_key
TRANSCRIPTION_GLOSSARY=Term1, Term2, Term3

3. Create the database

CREATE DATABASE IF NOT EXISTS srot_core;

Tables are created automatically on first startup.

---

Running

Terminal 1 - Backend:

source .venv/bin/activate
uvicorn backend.main:app --reload

API available at http://localhost:8000. Interactive docs at http://localhost:8000/docs.

Terminal 2 - Frontend:

source .venv/bin/activate
streamlit run frontend/app.py

UI available at http://localhost:8501.

---

Usage

Recorded Meetings

1. Go to the Upload page
2. Upload an audio file (mp3, wav, m4a, ogg, flac, webm)
3. The system runs the full intelligence pipeline in batch mode
4. Review the generated Hybrid Summary and extracted tasks
5. Edit task titles, descriptions, and assignees as needed
6. Click Create Jira Ticket to push to Jira and generate the Dev Task file

Live Meetings

1. Go to the Live page
2. Start a session; audio streams via WebSocket to the intelligence pipeline
3. Watch the transcript, context snapshots, and Jira suggestions update in real-time
4. After the meeting, the final transcript and summary are saved automatically

Code Indexing

1. Go to the Code Knowledge tab
2. Enter the project name and root path of your TypeScript/NestJS codebase
3. Click Index Project. The system parses all files, builds the Neo4j graph, and creates Qdrant embeddings
4. Define Domain Entities (business concepts like Calculator, MarksDistribution) and link them to services to improve extraction accuracy

---

API Reference

Meetings

Method	Endpoint	Description
POST	/process-meeting	Upload and process meeting audio (async)
GET	/meetings	List all meetings
GET	/meeting/{id}	Get transcript, tasks, and summary
PUT	/meeting/{id}/transcript	Update transcript and re-extract tasks
WS	/ws/intelligence/{id}	Streaming intelligence pipeline (PCM bytes)
POST	/start-live-meeting	Initialize a live meeting record
POST	/finalize-meeting/{id}	Finalize meeting status and extracted tasks

Tasks

Method	Endpoint	Description
PUT	/task/{id}	Update task title, description, or assignee
DELETE	/task/{id}	Dismiss a task
POST	/create-jira	Push a task to Jira and generate dev task file
GET	/dev-task	Fetch dev task file (?meeting=1&task=2 or ?jira=DEV-123)

Code Knowledge

Method	Endpoint	Description
POST	/index-project	Start background indexing of a codebase
GET	/index-jobs	List all indexing jobs
GET	/index-jobs/{id}	Get job status and progress
GET	/projects	List indexed projects
GET	/services/{project}	List services in a project
POST	/code-knowledge	Manually add code context (fallback)

Domain Entities

Method	Endpoint	Description
GET	/domain-entities	List domain entities with linked services
POST	/domain-entities	Create a domain entity
DELETE	/domain-entities/{id}	Delete a domain entity
POST	/domain-entities/link	Link a domain entity to a service in Neo4j

Related

Srot CLI connects your Srot task queue directly to Claude Code, launching an interactive AI session pre-loaded with your task requirements and project context.