Shreyansh Kumar SinghCaught in the blast
Survived 0.0s
I build AI products that hold up in production — RAG pipelines, LLM agents and autonomous security tools — on nearly four years of shipping .NET, SQL Server and Angular software at RamanByte. Below are nine of my projects, each with screenshots and a written case study.
How this page works — every project below is real, screenshotted live off a running server, not mocked up. Scroll and each one pins to the top of the stack as the next rises to cover it — like flipping through a stack of case files. No parallax library: the stacking itself is position: sticky, one per project; the only scroll-driven code left dims a card once the next one has covered it.
HOW I WORK
Before I write a single line of code, I ask:
Why does this need to exist?
Because a feature without a purpose is just another button.
Behind every click… there's a person trying to get something done.
So I don't design for screens.
I design for people.
An idea looks simple. Until you open the door.
My job is to make all that complexity feel like one simple experience.
What if the API dies? What if the user taps twice?
What if 100,000 people arrive tomorrow? What if everything goes wrong?
That's when the real engineering begins.
Software used to wait for instructions.
Now it can understand them.
That's where I believe software is going.
Nothing I build is ever really finished.
Every bug teaches. Every user teaches. Every failure teaches.
The next version is always waiting.
IN SHORT
I take an idea and turn it into software people can use — software that scales and survives real traffic, real data and real users.
AI-powered, full-stack software
that holds up in production.
Nine AI and full-stack projects by Shreyansh Kumar Singh — THE RANGE. Every repository below is real: the reference numbers are derived from commit dates (YYMM.CODE.serial), not invented. Keep scrolling — each project pins to the top of the stack as the next rises over it.
An AI-curated daily newsroom for trading cards, figures and luxury watches — with its own written design system.
An autonomous penetration-testing platform that validates every finding before it reports it.
An AI relationship-manager copilot for banking — RAG and document intelligence, then a written audit of its own P0 holes.
A walkable 3D museum where Indian art history sits as a first-class citizen on the world timeline.
Step inside any book — summaries, timelines, mind maps, character graphs and a grounded AI tutor.
Four physics simulators. One HTML file each. No build step, no dependencies, no network.
▶ Run it live — real canvas, this machine
Angular 18 and ASP.NET Core 8 in Clean Architecture — the enterprise half of the range.
An autonomous QA agent that re-verifies reported bugs against live application behavior — never against a commit message.
"Never guess." A bug is marked fixed only when observed application behavior says so — never because a commit exists.
A locked-off clip of a Lamborghini Revuelto, scrubbed frame-by-frame to scroll — scroll down and it assembles itself; scroll up and it comes apart again.
▶ Scroll it together yourself — opens in a new tab
Ten technical findings from six of my AI and full-stack projects — each one pinned under the project it came out of.
A thousandfold increase in yield multiplies the damage radius by ten, not a thousand — one exponent covers seven orders of magnitude.
Fallout dose becomes a function of position and time-since-detonation — activity falls ~100× in the first two days (Way–Wigner t^-1.2).
Every active check needs physical evidence — a DB error signature, a reflected payload, real command output — or it's discarded, not downgraded.
Internal-network scanning is a separate, explicitly authorized switch, host-locked with include/exclude rules and a rate limiter in front.
Embeddings, vectors, retrieval and generation are separate modules — each stage independently inspectable and replaceable.
P0 self-reported: no auth on any endpoint, sequential enumerable IDs — scrapeable with a for-loop. Published, not quietly patched.
Verified Wikipedia/Wikimedia data only; in-copyright modern artists become biography nodes rather than invented reproductions.
A focus raycast throttled to ~10×/s fixed perceived smoothness more than any framerate tuning did.
Three marques, three non-overlapping jobs — averaging references produces mud; assigning each a distinct role produces a system.
A bug is never marked fixed because code changed or a commit exists — only because its actual behavior in the QA environment, captured as evidence, says so.
How I build: write the specification before the code, then audit the result after it ships.
The shape of the thing gets written down first — what it is, what it looks like, where every number came from — so it can be picked up cold by anyone, including a future me.
Almost every repository ships a PROJECT_BIBLE.md and HANDOFF.md alongside the source, written to be picked up cold.
PROJECT_BIBLE · HANDOFF · AGENTS · CLAUDE — across 5 repositoriesThe visual language gets its own document with named surfaces and roles before a component exists. Reference the token, never hardcode a hex.
DESIGN_SYSTEM.md — every colour tokenisedBlast, thermal and fallout curves cite a published model. ANTARANG uses only verified Wikimedia data.
Model tables in README.mdThen it gets turned on itself — attacked like an outsider, run without its safety nets, and judged on whether the tool actually fit the job or just felt familiar.
SARTHI ships a full static self-review, graded P0 to P3, most damaging finding written first.
QA_AUDIT.mdBookVerse runs fully without an API key and says so with a badge. Nythera's internal mode is off by default.
Preview-mode fallback · internal mode default-offVanilla canvas physics. Python security tooling. R3F galleries. Next.js RAG. ASP.NET Clean Architecture. A canvas frame-sequence scrubbed to scroll instead of a video element. The range is deliberate.
9 repositories · TS · JS · Python · C# · WebGLSix working principles behind every system, workflow and decision above — how I approach AI and full-stack engineering.
AI stopped being a feature somewhere around the second project in this range. It's the layer other systems get built on top of — the one that turns a document pile into a bank's RAG tool, or nine findings into a research board. The advantage isn't access to a model anymore. It's knowing how to architect around one.
A one-off build stops paying you back the day it ships. A system — modules that specify, tokens that don't drift, numbers that trace to a source — keeps paying you back every time it's reused, extended or audited. Everything in THE RANGE is built to be picked up cold, including by a future version of me.
The parts of this work a model can't do are the parts worth protecting: deciding what a project is even for, choosing cube-root scaling over an ellipse, knowing which finding to report first. Automation raises the floor. It doesn't replace the judgment call at the top.
Execution got faster; taste didn't get automated with it. A spec that's actually legible, a design token that's actually reused, a self-audit that leads with the worst finding instead of burying it — that's the difference between work that reads as considered and work that reads as generated.
Nine repositories, five languages, one person — because the gap between having an idea and having it running got small enough to stop being a bottleneck. Building fast isn't a shortcut around rigor; the bible and the audit still get written. It's what makes that rigor affordable at this pace.
Nobody on this range shipped as "just a developer." Spec-writer, designer, security auditor, systems thinker, and the person deciding where the AI belongs in the loop — all the same person, on the same repository, in the same week. That convergence is the job now, not a side skill.
Direct answers to what recruiters, founders and collaborators ask most — who I am, what I've built, and how to work with me.
Shreyansh Kumar Singh is an AI and full-stack engineer based in Pune, India. He builds AI-native systems — RAG pipelines, LLM agent loops, autonomous security tooling — and has been a full-stack developer at RamanByte since January 2023, shipping production .NET, SQL Server and Angular software. This site, THE RANGE, is nine of his repositories stacked into one scroll.
Four AI systems where the model is load-bearing, not decorative: Sarthi, Nythera, HallogenAI and BookVerse AI. Sarthi is an AI relationship-manager copilot for banking — RAG and document intelligence — followed by a written audit of its own P0 holes. Nythera is an autonomous penetration-testing platform that validates every finding before it reports it. HallogenAI uses eight specialized agents to re-verify reported bugs against live application behavior. BookVerse AI turns any book into summaries, timelines, mind maps and a grounded tutor — with zero API keys required.
Yes — full-time or hybrid roles, based in Pune. His first choice is applied AI engineering (RAG, LLM agents); he's also a strong fit for full-stack engineering on AI products, AI security tooling, and senior .NET and Angular work, in that order of how strongly his shipped work supports them.
Email shreyanshkumarsingh208@gmail.com, or use the Contact page, which has one-tap call and WhatsApp buttons. The most useful first message covers three things: what's actually broken, what you've already tried, and what "done" looks like. Source code for the public projects is on GitHub at github.com/Shreyanskumarsingh20.
All 15 questions — experience, education, availability, stack →
The languages, frameworks and AI tools I use in production.