Argus- It never forgets

Argus is a proactive Enterprise Agent using Elasticsearch and Gemini 3 to turn messy WhatsApp chats into instant reminders. It eliminates "Information Silos" so you never forget.

Comment

Inspiration

We were inspired by a simple irony: we live in the age of Big Data, yet our most valuable personal data our conversations remains completely passive. It sits in a "Context Grave" while we make decisions in a browser. Think about it. You scroll past 200+ WhatsApp messages every day. Your friend mentions a restaurant. Your dad texts about your upcoming exams. You make a mental note to cancel your Netflix subscription. Then life happens you open Netflix, and you forget. You open an insurance form, and you type the wrong car year. You open a travel site, and you can't remember which hotel someone recommended three months ago. Mental notes don't have search bars. And you can't search for something you've already forgotten. We wanted to turn that "Passive Gold" into "Active Intelligence" an invisible layer that connects what you said last month to what you're doing right now. That's Argus.

(Personal note that hit home: one of us almost missed cancelling a Netflix subscription after exams, and the other misses quiz deadlines almost every time. We built the tool we desperately needed.)

What it does

Argus is a context-aware browser intelligence layer powered by Google Gemini, Elasticsearch Serverless, and Elastic Agent Builder. It silently watches your WhatsApp conversations and proactively surfaces the right information at the right moment — with zero manual input.

Key capabilities:

Proactive Context Popups: When you navigate to any website, Argus runs a hybrid semantic search across your WhatsApp history and slides in a gentle notification if something relevant exists. Open a travel site → reminded about a hotel your friend mentioned. Open Netflix → reminded about that subscription you planned to cancel.

Form Mismatch Detection: Argus reads live DOM form fields, cross-references them with your WhatsApp memory, and alerts you when something doesn't match. Open an insurance portal and accidentally type the wrong car year? Argus catches it and lets you fix it with one click — "Context-Aware Assistance with Human Approval."

Agentic AI Sidebar: Ask natural language questions like "Do I have any meetings this week?" and Argus enters a full agentic loop — Gemini decides what to search, calls Elastic Agent Builder MCP tools to query Elasticsearch directly, reads results, and can chain up to 5 tool calls before giving you a grounded, conversational answer . Time-Based Reminders: Polls Elasticsearch for due reminders at 24h, 1hr, and 15-minute pre-event windows and fires web-push notifications automatically — no calendar setup required. Three-Tier Resilience: Gemini goes down? Argus falls back to regex heuristics. Those fail? It pulls from an LRU response cache. Zero downtime.

How we built it

Backend: Node.js / TypeScript server orchestrating the full agentic pipeline. WhatsApp Integration: Real-time message ingestion via the Evolution API, feeding every incoming message directly into the processing pipeline. AI Layer: Google Gemini Flash powers intent understanding not keyword matching. It extracts events, reminders, deadlines, tasks, and recommendations from raw conversational text, then generates 768-dimension vector embeddings for semantic storage. Data Store: Elasticsearch Serverless (hosted on Elastic Cloud) with 6 purpose-built indices:

argus-events, argus-messages, argus-triggers, argus-contacts, argus-context-dismissals, argus-push-subscriptions

Search Engine: A hybrid kNN + BM25 pipeline merged via Reciprocal Rank Fusion (RRF):

kNN approximate nearest-neighbor search on the 768-dim embedding field (cosine similarity, k=10) BM25 full-text multi-match across title (3x boosted), keywords (2x), description, and location with AUTO fuzziness for typo tolerance

Agentic Orchestration: Elastic Agent Builder gives Gemini JSON-RPC tool access to query Elasticsearch directly, enabling multi-step reasoning loops. The pipeline includes a fallback orchestrator, background health-check agent, embedding backfill agent, time-based scheduler, and daily backup agent. Chrome Extension: Fires a context-check on every URL change, reads live DOM fields for form mismatch detection, renders notification cards, and hosts the AI sidebar

Accomplishments that we're proud of

Built a genuinely end-to-end agentic system from raw WhatsApp message to grounded browser notification with no manual user input at any step. Implemented hybrid semantic + keyword search with RRF on a real-world personal knowledge base, not a toy dataset. Achieved three-tier AI resilience Argus kept working flawlessly even during simulated Gemini outages. The form mismatch + human-approval flow is a novel UX pattern we're particularly proud of: AI that suggests, but never acts without your consent. The Netflix demo moment a real, genuine use case from our own lives working live on camera.

What we learned

Building Argus taught us that the hardest part of AI products isn't the AI it's the plumbing. Getting real-time data (WhatsApp messages) into a searchable, semantically rich store, and then surfacing it in the right context at the right moment, requires careful architecture across many layers.

We learned how powerful Elasticsearch's hybrid search capabilities are when combined with LLM-generated embeddings. The RRF fusion of kNN and BM25 consistently outperformed either method alone especially for short, conversational queries where pure semantic similarity sometimes drifts . Elastic Agent Builder fundamentally changed how we thought about AI tool use. Giving Gemini direct, structured access to query tools rather than stuffing everything into a prompt produces more reliable, grounded answers and enables genuine multi-step reasoning.

We also learned a lot about Argus itself specifically, how people's conversations contain far more actionable intelligence than anyone realizes. Once you start extracting and indexing this data, the patterns that emerge are remarkable: recurring commitments, implicit preferences, social obligations. The "Context Grave" is real, and it's full of gold.

Impact

Skills

Working on Argus forced growth across the full stack simultaneously — agentic AI orchestration, vector search architecture, real-time browser extension development, and resilient backend design all at once. You can't half-understand any of these; they have to work together.

Workflow

The core insight Argus proves is that your best workflow tool is already in your pocket — it's your chat history. Instead of duplicating information into todo apps, calendars, or note-taking tools, Argus makes the conversations you're already having do the work. It eliminates the "second inbox" problem where you have to manually migrate a WhatsApp reminder into another system.

Productivity

Argus removes the biggest invisible productivity tax: context-switching to go look for something you half-remember. The moment you land on a relevant page, the context comes to you. You never break flow to search your chats, and you never miss a commitment because you were in the wrong app at the wrong time.

What's next for Argus- It never forgets

Multi-platform ingestion: Expand beyond WhatsApp to Telegram, iMessage, email, and Slack building a truly unified personal context graph. Gemini Vision integration: Allow users to screenshot any content and have Argus extract and store events or context from images. Collaborative context: Let teams share relevant context with each other imagine Argus surfacing a colleague's WhatsApp note when you're both looking at the same client's webpage. Privacy-first on-device mode: Explore running the embedding and extraction pipeline locally for users with strict privacy requirements. Smart calendar sync: Auto-push extracted events to Google Calendar with one-tap approval, closing the loop between passive chat and active scheduling.

Share this project:

Updates