Inbox overflowing? Deadlines hiding in paragraphs? Yeah, we're not doing that anymore.
Meet ReplyLater AI — a Gmail assistant powered by Gemma 3 4B, running locally through Ollama.
It turns email chaos into structured, actionable insights.
AI Summaries — TL;DR, but actually useful.
Urgency Detection — High / Medium / Low.
Deadline Extraction — No more surprise due dates.
Action Items — Emails → actionable tasks.
Smart Replies — AI drafts, humans approve.
Google OAuth — Personal Gmail inboxes with separate user sessions.
https://drive.google.com/file/d/1MeajjvGwfE2NpGusfK38m-h90AZNy5xf/view?usp=sharing
https://github.com/arpitakashid08/replylater.git
How I Built It
01 / Connect. Integrated Gmail API with Google OAuth and separate user sessions.
02 / Infer. Used Ollama to run Gemma 3 locally, analysing selected emails without a paid model API.
03 / Structure. Enforced Pydantic-validated JSON with summaries, urgency, deadlines, action items, and suggested replies.
04 / Refine. Added a second-pass check when the model potentially misses deadlines or tasks.
05 / Ship the UI. Built a minimal React dashboard with an editable reply workflow. Nothing gets sent automatically.
Gmail → FastAPI → Gemma 3 (Ollama)
↓
Structured JSON
↓
Summary · Urgency · Tasks
Deadlines · Reply
Gemma gave me control over inference, prompts, output validation, and experimentation — all running locally.
No dependency on a proprietary inference API. No need to transmit email content to an external AI provider for local analysis.
Why Does Open Innovation Matter?
The core of ReplyLater AI is Gemma 3 4B, running through Ollama and integrated with a Python FastAPI backend.
My Agent Session
Here's how the AI pipeline works:
Ollama: Runs Gemma locally through its /api/chat endpoint.
Gemma 3: Analyses emails, detects urgency, extracts deadlines and action items, and drafts replies.
Structured Inference: Uses JSON schema constraints and Pydantic validation for consistent outputs.
Retry Logic: Triggers a second analysis pass when important deadlines or tasks might have been missed.
Privacy-Aware Workflow: Email analysis runs locally, without sending content to a third-party AI inference API.
The Lazy Payoff
A long email with three requests and a deadline buried in paragraph five?
Before: Read → Reread → Decode → Draft.
With ReplyLater AI: Select → Analyse → Review → Done.
No made-up productivity statistics. Just fewer repetitive steps between reading an email and knowing what to do.
Prize Categories
Best Use of Gemma 3 which powers the core email intelligence: summarization, classification, task extraction, and reply generation.
Top comments (0)