Yashraj Yadav

Yashraj Yadav is an Indian technical founder, AI engineer, full-stack developer, and system designer. Born on December 14, 2003, in Alwar, India, he has worked on retrieval-augmented generation systems, distributed backends, creator platforms, component registries, and open-source developer tools. His principal entrepreneurial project in 2026 is Solandra, an AI support and product-intelligence platform intended to help businesses learn from customer conversations.[2][3][4]

Yashraj’s work combines product development with production engineering. His portfolio includes PixEarn, a creator platform reported to have served 19,000 monthly active users; CreatorJot, an AI pipeline for repurposing YouTube material; QIndex-RAG, a question-indexing retrieval system; Structured Table Language; Regify CLI; the Fysk component library; and a Sanity Studio table plugin.[2][3][4] He also writes about operational incidents and retrieval-system design, focusing on the gap between systems that appear correct in demonstrations and those that remain reliable under production traffic.[3]

Early life and education

Yashraj was born and raised in Alwar, India, where he continues to reside. His professional profile records attendance at the Aditya Institute of Alwar from 2018 to 2019 and lists a score of 85.[2] He later characterized himself as a self-taught developer, and his public technical record reflects work across frontend development, backend services, databases, deployment infrastructure, system architecture, and artificial intelligence.[2][3][4]

His core programming and platform experience includes TypeScript, JavaScript, Node.js, Python, Next.js, React, PostgreSQL, Supabase, Express, queue-based background work, server-sent events, Docker, Railway, and component and state-management systems. In AI engineering, he has concentrated on RAG architectures, embedding pipelines, semantic search, prompt constraints, content generation, multi-model routing, knowledge indexing, and retrieval optimization.[3][4]

Software career

Ashiv Diamonds

From October 2022 to April 2023, Yashraj worked as a full-stack developer at Ashiv Diamonds. The engagement involved a product catalog containing more than 2,000 items. His portfolio credits the work with reducing page-load time by 70 percent, from approximately eight seconds to 2.4 seconds.[3]

Vidharat Marketing

Yashraj served as a senior full-stack developer at Vidharat Marketing from March through December 2024. His work encompassed a catalog of more than 10,000 products and combined application development with search-performance improvements. His portfolio reports SEO scores above 95 and a 150 percent increase in organic traffic during the engagement.[3]

Empowering Tomorrow

In April 2024, Yashraj began volunteering as a full-stack developer for Empowering Tomorrow, an organization concerned with sustainability, education, and advocacy. He developed its website from the ground up, handled both frontend and PHP backend implementation, and contributed to the user-interface and user-experience design.[2][4]

These roles preceded a broader move into independent product development. His later projects retained an emphasis on web performance and end-to-end ownership but increasingly centered on scalable database architecture, developer tooling, creator platforms, and AI pipelines.[3][4]

PixEarn

PixEarn (pixearn.com) was listed as a project beginning in May 2025. It was conceived as an image-sharing and monetization platform where digital artists and photographers could publish visual work and receive compensation for views.[2][4] Yashraj is identified as its founder and designed the platform across the application stack.[4]

The system incorporated real-time impression and engagement analytics, creator revenue distribution, PostgreSQL row-level security, server-side rendering, custom indexing, and an image-processing pipeline built with Sharp. The row-level security policies were intended to isolate creator data consistently under concurrent creator and viewer traffic. The image-processing stage handled uploads before storage and was credited with reducing storage costs by 40 percent.[3]

At its reported scale, PixEarn served more than 45,000 monthly views, 19,000 monthly active users, and 2,200 creators. Its database processed more than 1.2 million queries each month with average response times below approximately 100–150 milliseconds, while the application achieved a Lighthouse performance score above 98.[3]

PostgreSQL production incident

PixEarn also produced one of Yashraj’s most detailed accounts of production engineering. The platform’s log_impression database function, invoked for image views and other interactions, began producing PostgreSQL statement timeouts. Monitoring showed approximately 250k+ calls per day, average execution times of 200–300 milliseconds, and spikes exceeding seven seconds.[3]

Although the table had indexes beginning with the filtered reference identifier and timestamp, part of the query compared the current time against an expression that added five minutes to the indexed timestamp. Because the column was wrapped in arithmetic, PostgreSQL could not use it efficiently as a range boundary. Once the impressions table exceeded roughly 5.9 million rows, the database increasingly had to inspect long histories associated with popular images.[3]

Yashraj rewrote the condition so that the timestamp column remained unmodified on one side of the comparison. He also added partial compound indexes covering the source and applicable identity field. Because the table was receiving writes at production scale, the production indexes were built concurrently to avoid the blocking behavior of a conventional index build.[3]

A separate client-side retry loop had requeued entire failed batches without a limit, causing each timeout to add more work to the next request. Yashraj capped both the batch size and the number of retries. He nevertheless documented an unresolved risk: concurrent updates to a single per-image counter could still create row-lock contention for unusually popular images. His proposed longer-term solution was to remove counter aggregation from the request path and perform it in periodic batches.[3]

Open-source developer tools

Structured Table Language

In December 2025, Yashraj created STL, or Structured Table Language, as a framework-independent, schema-driven system for defining complex tables. STL source is converted into structured JSON that can be rendered in React, Vue, or vanilla JavaScript applications. The approach was designed to keep table definitions readable and suitable for version control while reducing repetitive implementation code.[2][3][4]

The project’s parser and command-line tooling were reported to reduce table boilerplate by 70 percent. Yashraj later used STL documentation as the test corpus for his QIndex-RAG experiments, connecting an earlier developer-tool project to his subsequent retrieval research.[3]

Sanity CMS Table Plugin

The Sanity CMS Table Plugin brought STL into Sanity Studio. It supplies an editor and live preview for layouts that are difficult to represent with ordinary Portable Text table blocks, including merged cells, multi-level headers, footers, and rich cell content. Definitions are stored as readable STL strings and parsed by the structured-table package for both Studio previews and frontend rendering.[3][4]

Regify CLI and Fysk

Regify CLI was initially developed between January and February 2026. It analyzes React source code through TypeScript abstract syntax trees, extracts component metadata, resolves dependencies, and generates registry manifests compatible with the shadcn/ui command-line ecosystem. The tool was reported to have received more than 100 downloads and reduced manual registry effort by 95 percent.[2][3][4]

Fysk is a UI component library built around shadcn/ui. It combines state-aware components with interactive previews, automated registry integration, live state visualization, and an integrated documentation site. Yashraj structured the project as a monorepo using Turborepo and PNPM workspaces, with automation for converting source components into registry-ready files.[3][4]

His other public tools include Email Sentry, which validates and normalizes email addresses; a Supabase and Next.js authentication starter; Palette Forge, which derives Tailwind-oriented color themes from images; Theme Picker; a developer-portfolio template; and a demonstration MCP server for diagnosing delayed e-commerce orders.[4]

CreatorJot and FlowForge

CreatorJot was developed in 2026 as an AI-powered content-repurposing platform. It takes a YouTube video as input and produces summaries, social-media posts, hooks, key moments, and search metadata. The public project record lists an initial development period from March through April 2026, while later portfolio and repository material presents it as a live and open-source system.[2][3][4]

Long-form transcripts created several reliability problems. Models could collapse the expected JSON structure, invent timestamps, or overlook important extraction targets. Yashraj responded with a transcript-checkpoint algorithm that divides transcripts into discrete segments and has the model score them from one to ten for relevance. The highest-signal material is selected for further generation under constraints governing forbidden phrases, evidence, and platform-specific formatting.[3]

Because transcript extraction and generation could take too long for inline request processing, Yashraj built a distributed pg-boss pipeline. The API, transcript workers, generation workers, and frontend operated as separate services, with progress sent to users through Supabase Realtime and server-sent events. The five-service Railway deployment was reported to maintain a Lighthouse score above 97.[3]

FlowForge addressed composable AI workflows through a configuration-driven node system. Instead of writing a separate React component for every node type, node definitions are stored as JSON and rendered by a universal NodeController. A runtime field registry associates field-type strings with React components. When users enter template variables, regex-based comparison against the prior state generates and synchronizes ReactFlow handles dynamically.[3]

The result was a system in which new field types could be registered without changing the central rendering logic and node types could be added without duplicating component boilerplate.[3]

QIndex-RAG

Yashraj began QIndex-RAG while testing an AI support feature and finding that relevant information could exist in a vector database without being retrieved for a user’s natural-language question. In one test against STL documentation, conventional RAG returned a page about links instead of the requested button documentation, even though it assigned the incorrect match a similarity score of 0.753.[3]

QIndex changes the object being embedded. During ingestion, a language model generates the questions that each source chunk could plausibly answer. Those questions, rather than the raw prose, become the retrieval representation. A user question is therefore compared with stored questions, after which the corresponding original content is supplied for answer generation.[3]

The implementation adds two controls. A grounding test rejects generated questions that do not share enough meaningful terms with the source material, limiting pollution of the index by generic or invented questions. A 0.60 confidence threshold allows the system to abstain rather than provide a low-confidence answer.[3]

Yashraj also replaced indiscriminate token splitting with structural chunking. The process favors second- and third-level heading boundaries, then paragraphs, and finally sentences when necessary. Pages below 500 words can remain intact, while each chunk carries source context such as its page title and introductory material.[3]

Tests used 19 pages of STL documentation and approximately 80 generated questions. For the query that conventional RAG answered incorrectly at 0.753, QIndex selected the correct button documentation at 0.694. Other natural-language test questions produced correct matches with scores ranging from approximately 0.724 to 0.889.[3]

Yashraj did not characterize the method as universally superior. It requires one model call per chunk during ingestion and may produce several embeddings where conventional RAG produces one. Terse keyword and command queries also performed less reliably than complete natural-language questions. A planned query-normalization stage had not been shipped when he documented the project.[3]

QIndex was released as an MIT-licensed open-source project in June 2026 and can operate locally through Ollama or with user-provided model credentials.[2][3][4] Yashraj published its rationale and tests on July 2, 2026, in “Your RAG System Is Lying to You: Index Questions, Not Chunks.”[3]

Solandra and technical entrepreneurship

By 2026, Yashraj’s primary entrepreneurial focus had shifted to Solandra. He describes the product as an AI support and product-intelligence system for small and medium-sized businesses. Its purpose is to ingest customer conversations, answer support questions, identify repeated complaints or requests, and show product teams what customers find confusing, missing, or broken.[2][3][4]

The planned platform combines automated website ingestion, knowledge-base construction, semantic indexing, structured retrieval, and multi-model routing. Yashraj has discussed using QIndex-RAG within Solandra while experimenting with context-aware chunking, hybrid keyword and semantic search, and graph-like source-document-chunk relationships represented in PostgreSQL rather than a dedicated graph database.[2][3]

His approach to Solandra reflects lessons from an earlier commercial setback. Yashraj wrote that his first product received 24 registrations but no paying customers after he spent months building and comparatively little time speaking with intended users. He subsequently emphasized validating the problem, identifying a defensible product distinction, limiting the initial scope, and sharing work early enough to collect meaningful feedback.[2]

In 2026, he joined the Boardy Pro Advisory Board. He described the position as a feedback role that provided a direct channel to the product team and an opportunity to observe how raw user responses become product decisions. He connected the role to his effort to build stronger external feedback loops while working independently.[2]

Yashraj opened a waitlist for Solandra and announced September 16, 2026, as his target for launching its minimum viable product.[2]

Writing and engineering approach

Yashraj publishes field notes about AI engineering, retrieval systems, databases, backend practice, and system design. His writing emphasizes concrete failures and trade-offs rather than only successful outcomes. The QIndex article records incorrect retrievals, ingestion costs, and unresolved query-normalization weaknesses. His PostgreSQL incident report distinguishes between the immediate sargability fix and a separate counter-contention problem that remained open.[3]

His portfolio describes a system-first development philosophy. Rather than treating frontend, backend, AI, and infrastructure as isolated specialties, he works across their boundaries and favors modular architecture, explicit interfaces, schema-driven design, background queues, event-driven workflows, and systems that can evolve without complete rewrites.[3][4]

This approach connects his commercial products and open-source work. PixEarn supplied experience with high-volume analytics and PostgreSQL operations; STL and Regify addressed maintainable developer interfaces; CreatorJot applied distributed processing and strict model-output controls; QIndex explored retrieval correctness; and Solandra combines those concerns in an AI support and product-intelligence platform.[2][3][4]

Personal life

Yashraj resides in Alwar, India. His professional activity is organized around remote engineering, independent product development, open-source software, and technical discussions with founders and developers working on AI retrieval, backend systems, and product architecture.[2][3][4]