Open to architecture, AI platform and backend engagements
Fig. 00 — The desk, after the 2022 room
Solution Designer & AI/Platform Engineer
I design and ship production systems end to end: applied AI, event-driven backends and secure platforms that clients sign off on.
Illustration: a paper diorama of Nischay’s desk. On the wall hang a window, a clock set to Surat time and a pinned sketch of a RAG pipeline. On the desk sit a laptop showing 148 of 148 tests passing, a notebook with “22 to 0” circled, an index card reading 149x, a mug and a lamp that switches on in the dark theme.
3+ years shipping production systems, now owning solution design across concurrent enterprise client platforms — from problem framing and architecture through prototype, hardening, production rollout and client UAT sign-off. Previously Lead Developer for a full product team.
If I say it works, there’s a number behind it. Each one opens the system it came from.
Each project shows how the system thinks: the architecture, the hard decisions and the numbers that came out of it.
Turns code-mixed voice notes from the field into governed, queryable business intelligence — with every model running on local GPU.
One unified telemetry stream from many vendors, powering a platform that runs an electric delivery fleet end to end.
Medical-education events tracked from money down to evidence of delivery.
Raw social data → embeddings → communities, explored as an interactive influence graph.
§ 03 · The platform, in motion
One island, drawn from parts of two client platforms. Take the tour and the camera follows a single scooter reading in, then a single question all the way to its answer.
A scooter reports in
A delivery scooter's battery and position reach its vendor, and the vendor calls my webhook. The gateway checks the hashed bearer token before anything else happens.
From · EV fleet telematicsCase study
Scroll to followClick any station to open its project
A delivery scooter's battery and position reach its vendor, and the vendor calls my webhook. The gateway checks the hashed bearer token before anything else happens.
The raw payload is saved, then batched onto AWS SQS. A pool of workers drains the queue. A transient failure lands in the dead-letter queue and is retried automatically.
The normaliser maps every vendor's fields and CAN codes onto one schema, converts time to UTC and keeps unmapped keys. It writes de-duplicated history and a live snapshot that moves the dot on the ops map.
Later, a regional manager asks what the field team heard this week. Microsoft Entra SSO signs them in with OIDC and PKCE, and the API issues a short-lived token carrying their scope.
API scope, then Postgres row-level security, then a scope filter inside the vector query itself. Each layer only lets through what this manager is allowed to see.
Vector search finds the nearest neighbours while Postgres full-text matches the words. Rank fusion merges them, a cross-encoder re-sorts them, and a local GPU writes the answer. Nothing leaves the network.
The answer goes back with inline citations, and an append-only audit row is written. That's the whole system: every part of it is running for a real client.
Composite illustration: telemetry intake, the queue and dead-letter handling come from the EV fleet platform; sign-in, access layers, retrieval, reranking and local GPU inference come from the field feedback platform.
The question from chapter four, slowed down: how the RAG assistant I designed for a global consumer-health client decides what each question deserves. SQL, retrieval or both.
A manager asks something countable and something human. Most assistants send both to a language model. This one reads the question first.
Statistical, trend and location questions are routed to SQL over pre-computed TimescaleDB rollups. The answer is exact, instant and costs zero LLM calls.
Qualitative questions run hybrid retrieval: vector search and Postgres full-text search in parallel, fused with Reciprocal Rank Fusion. Access scope is injected into the vector query itself, so out-of-scope feedback is never retrieved.
A cross-encoder re-reads each candidate against the question and reorders them. What the embedding thought was close isn't always what's useful.
Generation is grounded in the top passages with inline citations, capped at two LLM calls per answer, and every query writes an append-only audit row.
Illustrative questions · real mechanism
From AI services, to leading a product team, to owning solution design across enterprise platforms.
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.
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.
Software Engineer — AI & Platform (Full Stack)
Built AI-heavy platform services across .NET, Python and Node.
Grouped the way I think about a system. Switch the lens at the top of the page and this index re-orders itself.
Ordered for AI platform
Where the habits came from: an LLM scoring layer, an identity platform and two realtime apps that are live on Google Play.