Avash Shrestha

The problem

Luxury hotels sell packages and experiences on their own websites, and those offers change constantly. Entering them by hand took weeks per cycle, and some offers expired before anyone could type them in.

What I built

• A Python and Celery pipeline that crawled 1,200+ hotel properties across 130+ domains and used the Gemini API to extract structured offers.
• Multi-stage extraction: content-quality gating, page-type classification, traversal from list pages into each offer's detail page, schema-constrained prompting, and staged validation against hallucinated output.
• A second Gemini stage that scored every offer on six quality metrics, against rubrics defined with product stakeholders, citing the evidence behind each score.
• The React and TypeScript operations platform where the team managed, cloned, and test-ran 1,950+ crawl configurations, tracked live crawl progress, and reviewed extracted offers against a markdown view of each source page.
• A re-architecture of the single-stage Django/Celery pipeline into fetch and process stages on separate workers.
• A rebuilt Django ingestion for the vendor feed of hotel property data, hundreds of thousands of PostgreSQL records, using bulk writes and removing N+1 queries from the main write paths. A Celery handover passed an overrunning sync's unprocessed IDs to the next run, so update bursts of 8,000+ properties (8x the vendor's stated ceiling) no longer stacked up parallel syncs that blocked other tasks.
• A written make-vs-buy analysis that moved crawling to two scraping vendors, plus a provider-neutral fetch layer over them, hosted on DigitalOcean and Cloudflare.

The problem

An agricultural labor client tracked crew time and work on a legacy Java Android app with no documentation, and its author was retiring.

Crews worked at job sites with poor connectivity, and their records fed payroll, billing, and wage-and-hour compliance for the client's customers, so data captured in the field could not be lost.

What I built

• Worked out the undocumented app's behavior and rebuilt it in React Native and TypeScript: 21 screens covering crew login, time tracking, work capture, signatures, and reporting in one workflow, on React Native 0.79 and the New Architecture.
• Hardware capture for NFC/RFID tags, HID barcode scanners, and USB weight scales, with a custom Kotlin native module for serial scale reads and a key-event pipeline that deduplicates scans.
• Offline-first storage in Realm: per-day event logs, 500-event batch uploads on a 5-second interval, and a Kotlin storage module that falls back to internal storage and re-queues failed writes.
• English and Spanish support through a small localization engine built on React context, plus dark mode and tablet layouts readable in direct sunlight.

The problem

The US division of a global agriculture company used this platform to reach growers, retailers, and consultants. Agronomists and the client's staff published updates, market analysis, and field conditions across an iOS and Android app, a website, and an admin site.

Partway through a major rebuild, our tech lead left. The release date stayed the same, and there was no backfill hire.

What I built

• Took over the tech lead's mobile and DevOps work, led the frontend, and re-scoped the release into phases the client prioritized.
• Rebuilt the React Native app for iOS and Android and revamped its companion React website in about 3 months.
• Chat and event livestreams in React and React Native on the Stream SDK, with push notifications that opened the right screen.
• A hybrid update strategy: CodePush for JavaScript fixes, Firebase Remote Config and feature flags for runtime changes, and forced-update prompts for native releases.
• Branch-triggered CI/CD with Fastlane and Bitrise for iOS and Android, and CircleCI for the React and Next.js web deploys.
• Co-led a visual ROI calculator comparing the client's grower rebates with three competitors', which its sales and marketing teams used with growers.

The problem

Agricultural market users needed to act on time-sensitive prices. At the same time, each of our products needed the same chat, CRM, media, and login features.

What I built

• Dashboards and a futures board in React and React Native over live market data, comparing multiple time series across configurable date ranges.
• Django authentication flows, including a login-as-user path that let support troubleshoot client accounts without exposing unauthorized actions.
• Shared package libraries for CRM, chat, media, authentication, and core UI primitives across React, React Native, and Django, with TypeScript-typed interfaces.
• Upgraded multiple Django backends from version 2 to 3 on my own.