Llamail

I created a self-hosted email agent that combines hybrid retrieval with a small local model to search my mailbox, answer questions with citations, draft replies, and route actions through Telegram.

I designed and implemented the FastAPI services, routing, retrieval, local-model integration, storage, Telegram workflow, and operator interface.

Read the implementation series →All selected work →

Sanitized Sable Telegram UI · cited answer and operator-visible email analysis

Role: Independent system · implemented end to end
Runtime: FastAPI · llama.cpp · local GPU
Retrieval: ChromaDB + SQLite FTS5
Interface: Telegram via n8n

01 Delivery

How a request moves through Llamail.

Slash commands and recognized compound commands go directly to application handlers. For free-form requests, the local model classifies the intent, then application code validates the result and calls the matching handler.

Sable is the Telegram interface and operator persona for Llamail. It presents cited answers, reports incoming-email analysis, and shows drafts or mailbox actions for review before execution.

Command routing

Known slash and compound commands go straight to handlers. Free-form messages return a structured intent from the local classifier to the same validated application handlers.

Hybrid retrieval with citations

The retriever combines semantic search in ChromaDB with keyword search in SQLite FTS5, and answers link back to the retrieved email evidence.

Human-reviewed actions

The interface shows the interpreted request and proposed draft or mailbox action for operator review before execution.

Measured on local hardware

On my RTX 4080 Mobile 12 GB system, hybrid search took about 3 seconds, retrieval plus answer generation about 7 seconds, and natural-language intent classification about 4 seconds.

Technical writing

Read the five-part Llamail implementation series.

From Inbox to Character: Creating a Private, Local AI Email Agent →
The complete system, local-model runtime, Telegram interface, storage, and operating boundary.
How /search and /ask Work: Local Hybrid RAG with ChromaDB + SQLite FTS5 →
Weighted semantic and keyword retrieval across a large mailbox, with relevance-aware citations.
LLM as Router: Intent Classification for a Local Telegram Email Agent →
Slash commands and deterministic handlers first; natural-language classification only as a fallback.
Implementation Lessons: Hidden Headaches of a Local Gmail AI Agent →
Gmail synchronization, labels, threads, identifiers, and the failure modes that shaped the design.
LLM-Personalized Email Campaigns with Python, Gmail, and Local AI →
Previewed personalization, throttled delivery, reply tracking, and conservative classification.

02 Interface evidence

Telegram interface and system architecture.

The sanitized Telegram capture shows cited answers and incoming-email analysis. The diagram maps Gmail and Telegram onto the local FastAPI, retrieval, storage, and model services.

Llamail architecture diagram separating Telegram and Gmail services from the local FastAPI, retrieval, storage, and model runtime — System boundary · external Gmail and Telegram routes around local inference, retrieval, and state

03 Delivery boundary

Where Llamail runs.

Local inference, embeddings, FastAPI services, SQLite FTS5, and ChromaDB run on my machine. Gmail and Telegram provide the external mailbox and messaging connections through their APIs.
Running Llamail requires user-owned Gmail and Telegram integrations plus local model hardware, so this project page uses sanitized screenshots in place of a public demo.