About · Surat, Gujarat, India
I like the moment a diagram survives production.
I'm Nischay, a solution designer and engineer based in Surat, India. I studied Information Technology at SCET from 2020 to 2024, and joined Xtreme Thoughts SoftTech in January 2024 to build AI and platform services.
Since then I've worked on every part of a system's life: framing the problem with a client, drawing the architecture, writing the first prototype, hardening it, rolling it out and sitting through UAT until it's signed off. I've led a product team at Shamonk, and I now own solution design across several enterprise platforms at once.
Most of what I build is for clients I can't name: a global consumer-health company, a pharmaceutical firm, an electric delivery fleet. So this site shows how my systems think rather than what their screens look like. Every diagram here is a simplified redrawing, and every number is real.
Six habits, each with a receipt
Not values on a wall. Each one comes from a specific system and a specific decision.
Take the cheapest correct path first
If SQL can answer a question, a language model shouldn't. In the field feedback platform, statistical questions run on rollups at zero LLM calls, and only the genuinely qualitative ones reach retrieval.
Field Feedback Intelligence & RAG Analytics Platform →Make retries boring
Pipelines should be safe to run twice. Each stage clears what it owns before rewriting, failures land in a dead-letter queue with a reason, and a retry is a button, not an incident.
EV Fleet Operations Platform & Telematics Ingestion →Security in layers, not in hope
Access is checked at the API, again in Postgres row-level security, and again inside the vector query. If one layer forgets, the next one doesn't.
Field Feedback Intelligence & RAG Analytics Platform →Change shouldn't need a deploy
New telematics providers are onboarded from an admin screen, and the cutover from vendor APIs sat behind a feature flag so it could be undone in a click.
EV Fleet Operations Platform & Telematics Ingestion →Measure the claim
148 of 148 backend tests and 106 client UAT cases before launch. 97% agreement between the LLM scorer and human reviewers. If I say it works, there's a number behind it.
AI Audio Analytics & Facial Recognition →Follow the bug down to the wire
A two-second delay on every model call turned out to be IPv6 name resolution. Fixing it made vector queries 149 times faster.
Field Feedback Intelligence & RAG Analytics Platform →
Where I’ve worked
- Nov 2025 — PresentNowSurat, India
Xtreme Thoughts SoftTech
Solution Designer & Senior Engineer
Also Embedded with Connexink as Solutions Architect & Senior Engineer, for their outsourced client accounts.
Returned to Xtreme Thoughts to lead the AI engineering track and own solution design across concurrent enterprise platforms, from problem framing and architecture through prototype, hardening, production rollout and client UAT. Also embedded with Connexink as Solutions Architect for their outsourced client accounts.
- Architected a field feedback intelligence platform for a global consumer-health client: unstructured text and multi-speaker Hindi/English voice recordings become RBAC-scoped business intelligence for a five-tier sales hierarchy, with all inference on local GPU.
- Built a hybrid RAG assistant behind a question-type router — statistical questions hit TimescaleDB rollups at zero LLM cost; qualitative ones run vector + full-text retrieval with RRF, cross-encoder rerank and grounded citations, capped at two LLM calls per answer.
- Enforced RBAC in three independent layers: API scope dependencies, Postgres row-level security via per-session GUC, and scope filters injected into the vector query.
- Closed a self-improving vocabulary loop (HDBSCAN drift detection, schema-constrained LLM induction, semantic guardrails) that took the review queue from 22 near-duplicate proposals to effectively zero.
- Diagnosed a native-library load-order fault silently killing workers and an IPv6 resolution issue adding 2s to every model call — 149x faster vector queries.
- Architected telematics ingestion for an EV fleet platform (webhooks → SQS → unified schema, DLQ auto-retry, per-provider health) — shipped with 148/148 backend and 106 client UAT cases passing.
- Delivered an enterprise event & budget lifecycle portal for a pharmaceutical client on ASP.NET Core 8 + Dapper + SQL Server with Azure AD SSO.
- Built a data intelligence & network visualisation platform turning unstructured social data into embeddings, semantic clusters and an interactive influence graph.
- Feb 2025 — Oct 2025Surat, India
Shamonk Pvt. Ltd.
Full Stack Engineer & Lead Developer
Owned product engineering end to end for a consumer marketplace connecting users with expert agents, and set the team's technical roadmap.
- Architected a User Management System from scratch on Spring Boot serving millions of API requests, with fine-grained role & permission authorization across hundreds of endpoints.
- Designed and shipped two production React Native apps (User and Agent) with real-time chat and voice/video over WebRTC on custom signalling, scaling to 10,000+ concurrent users.
- Ran infrastructure on AWS and GCP over PostgreSQL, MongoDB and Redis at a 99.9% uptime target; Jenkins/Docker pipelines and Prometheus monitoring cut manual deployment effort by 60%.
- Jan 2024 — Jan 2025Surat, India
Xtreme Thoughts SoftTech
Software Engineer — AI & Platform (Full Stack)
Built AI-heavy platform services across .NET, Python and Node.
- Built an AI audio analytics platform in C#/.NET analysing millions of support calls at ~2x previous throughput, with an LLM evaluation layer scoring agents at 97% agreement with human reviewers.
- Delivered a facial recognition engine at 98% accuracy, matching 5,000–10,000 images in under 10 seconds via an optimised OpenCV + CUDA embedding pipeline.
- Shipped an AI form-processing service classifying and routing 100,000+ submissions with automated email workflows (Python + NestJS).
- Automated 200-page report generation in 20 seconds with Puppeteer, cutting report storage by 80%.
Off the clock
Some things don't change
The 2022 version of this site had a parallax room I drew layer by layer, a Luffy quote and a Zoro favicon. The room grew up into the desk on the home page. The crew is still around, if you know where to look.