Backend engineer building Sift, a personalized news app. Python, FastAPI, data systems, AI, and product building.

Kumar Sagar Maiti • Peerlist

<h3>Personalized News for Signal-First Intelligence</h3>Sift offers a public beta for a new approach to news consumption, focusing on what truly matters to you and filtering out the noise. Instead of overwhelming you with broad headlines or noisy social feeds, Sift provides condensed, actionable briefings on the companies, people, policies, sectors, and events you choose to follow.<ul><li>Entity Tracking: Follow specific entities like 'OpenAI' or 'Nvidia', not just general categories like 'AI'.</li><li>Event Collapse: Related news stories are grouped into single updates, eliminating redundant headlines from multiple sources.</li><li>Personalized Ranking: Your preferences, boosts, and suppressions influence what appears, ensuring the feed adapts to your specific interests.</li><li>Briefings, Not Doomscrolling: Designed for weekly intelligence consumption, allowing you to stay informed with a quick read, rather than getting lost in endless feeds.</li></ul>The core problem Sift addresses is the overwhelming volume of information. Sift's workflow is designed for simplicity: you define what matters, Sift monitors changes, and you receive concise briefings explaining what changed, why it's important, and what to anticipate next.Who it's for:<ul><li>Founders tracking competitors and policy changes.</li><li>Builders focused on AI and startup ecosystems.</li><li>Investors monitoring companies and market sectors.</li><li>Researchers analyzing policies and institutions.</li><li>Operators who are keeping up with market and regulatory shifts.</li><li>News readers who are tired of the volume of irrelevant news.</li></ul>Sift is currently in a public beta. Download the public beta from the Play Store to experience a more focused and intelligent way to stay informed.

Sift — Personalized News

Owners can register their businesses and movies. Users can search for movies, and halls near them showing the film and buy tickets for the same. While booking is under process no other user can book those selected seats. But if the order is not made within 5 mins the seats become available again. This was achieved using MongoDB TTL Indexing. I have also used authentication, authorisation and Axios to fetch movies from open-source OmdbAPI.

Movie Ticket Booking Backend

In this project, we created APIs for e-commerce backends. We created user, products, cart and order models and APIs. We implemented AWS and stored images as S3 links. Our database was MongoDB.

Shopping Cart

Made APIs for shortening long URLs and fetching them from database and cache memory. Implemented caching with Redis.

URL Shortener

Created APIs for new users, books and reviews. Users can also update, edit or delete their books and reviews. Guest reviews and ratings are stored as well. All the data are stored in MongoDB, created and managed with the mongoose library. Handled authentication and proper authorisation with expiry token.

Books Management

Built APIs for creating new authors and blogs in MongoDB. Managing to edit, fetching and deleting blog data with NodeJS and express. Provided authentication and authorisation with JSON web token handling all the edge cases.

Mini Blog Management

Managed APIs for creating new interns and new college data. Stored them in MongoDB. Implemented project with frontend code and managed multi-form data and handled CORS errors.

Open to Interns

Entity-driven news intelligence app — built solo, production-running. Ingestion pipeline processes 49,800+ articles across a knowledge graph of 130,000+ entities. Redesigned article processing architecture to improve pipeline reliability from 71% to 95.7%. Cut AI inference costs by 75% (₹500–1,000/day → ₹100–250/day) by migrating 98% of extraction workload to DeepSeek with Gemini as automatic fallback. Key systems built from scratch: — Multi-model LLM orchestration with cost-aware routing — pgvector semantic deduplication with entity-overlap + headline Jaccard fallback — Non-destructive identity resolution layer (canonical entity grouping across 130K+ entities without rewriting foreign keys) — Alias contamination detection to reject LLM-injected co-reference noise

Sift - Personalised News - siftnews.in

Open-source multi-model deliberation platform — forked and extended from Andrej Karpathy's LLM Council. MIT licensed. Core problem: frontier model deliberation required paid API keys by default. Built hybrid council mode mixing local Ollama models with optional cloud APIs (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter) — zero mandatory cost to run. Built on top of the fork: — Anonymised peer-review protocol (A/B/C labelling) to eliminate model-name bias in cross-model critique and chairman synthesis — Hardware-aware Model Manager with local model recommendations, install/uninstall controls, and availability checking — Hybrid council mode: mix local + cloud models in single deliberation session Stack: Python · FastAPI · React · Vite · Ollama Open for contributions.