Gurman SinghFull stackCloudAI systems
I’m a full-stack engineer and AWS Certified Solutions Architect. I take complete ownership of your product, from the interface your customers use to the cloud, queues and AI behind it, and I keep it fast, secure and running long after launch.
Taking on new projects · replies within one business day
AWS Certified
Solutions Architect – Associate
Lighthouse · mobile● measured
$ ollama serve
GET article:812HIT
Built and running for
§ 01What I own for you
Projects usually stall in the gaps between a designer, a developer, a DevOps contractor and an AI consultant. I cover all five layers myself, so nothing falls between them and you always know who is responsible.
Also handled
§ 02Selected work
Every one is live and in daily use. Each case study shows what the product is, what I built, the evidence behind it, and what it means for a business like yours.
01 / 06Productivity PaaS · Flagship
A calm home for work,clients and money.
FIG. 1.1The live Margin home screen: today’s work, the day log, habits and everything waiting on you, in one view.
Freelancers and small teams usually run their week across five subscriptions. Margin replaces them with one keyboard-first workspace: tasks, projects, clients, money, messages, a journal and habits, sharing one design, one search, and a password vault the server cannot read.
Role
Product, design & full-stack build
Stack
Next.js · React · Node.js · Tailwind CSS
Scope
14 tools, encrypted vault, messaging & calls
Each tool would be a product on its own. In Margin they share components, data and a single command palette, so a payment marked paid lands in Finance and a meeting’s action items become tasks.
Tasks
Day-by-day to-dos with priorities, tags, multi-day tasks and a daily routine.
Passwords are encrypted in the browser with a key derived from the master password. What reaches the server is ciphertext, so even a full database leak reveals nothing.
Your browser
Master password
Derive key
Plaintext
Margin server
Stored
Server can read it?
FIG. 1.2Zero-knowledge storage. The key exists only on your device; the server stores and syncs what it cannot read.
02 / 06Publishing platform · AI
Where Indian writersfind their readers.
FIG. 2.1The AI score flags what makes writing read as pasted from a chatbot, then fixes most of it in one click.
A free publishing platform for independent writers in India, in English and Hindi. The editor is the product: it reviews drafts, scores them for chatbot habits, screens content, checks SEO, then publishes a fast, server-rendered page that search engines can find.
Role
Product, design & full-stack build
Stack
Next.js (SSR) · Node.js · RTK Query · OpenAI APIs
Scope
AI editor, moderation, SEO, scheduling, versions
Live
Six checks between a blank page and a ranked article. Writers keep control: every AI suggestion is optional, and nothing changes what the story says.
01
Draft
Distraction-free editor with autosave and versioning
How AI search is changing SEO
38 words · saved
02
AI review
Marks unclear sentences and tone, one-click accept
Clarity 58 → 79
3 sentences sharpened
03
AI score
Scores 0–100 for chatbot tells and fixes most in one click
AI score 100 → 6
8 patterns fixed
04
Moderation
Flags spam and unsafe content before anything goes live
Content check
Clear ✓
05
SEO check
Title, summary, headings, keyword and readability
SEO score 46 → 92
Ready to rank
06
Publish
Server-rendered page, sitemap entry, now or scheduled
Live · indexed
publisha.in/articles/…
FIG. 2.2The publishing pipeline, step by step. Readouts show an illustrative run through the real editor’s checks.
01
Every save is kept with word-level diffs; restoring keeps the URL and live status.
02
Queue a story up to 14 days ahead and it goes live on its own.
03
Server-rendered article pages, structured metadata and a sitemap search engines read.
03 & 04 / 06Premium brand sites · Australia
Two businesses where the website is the first impression. Both are photo-led, both still load fast on a mid-range phone, and both score 100 for accessibility, best practices and SEO.
03 / 06 · Canberra
Architect-led custom homes, knockdown rebuilds and duplexes. Building across Canberra since 2012.
Lighthouse · deolbuilt.com.au
04 / 06 · Canberra
A calm family dental practice in Dickson, Canberra, built around unhurried care and easy booking.
Lighthouse · dicksondental.com.au
FIG. 3.1Lighthouse 12, measured 3 November 2026 with default mobile throttling and the desktop preset. Mobile performance moves a few points between runs; the other three categories hold at 100.
05 / 06Medical book marketplace · India
India’s largest onlinemedical bookstore.
A marketplace for medical students and professionals: a 45,000-title catalogue sold through a website and Android and iOS apps, with separate portals for sellers and affiliates, quick dispatch and international shipping.
Buyers
Web · Android · iOS
Sellers
Seller portal
Affiliates
Affiliate portal
One platform
FIG. 4.1Three audiences on one catalogue, order and fulfilment system.
06 / 06Custom concrete · Greater Toronto
Premium concrete,choreographed.
A premium contractor needed a site that feels as considered as the work. The answer was motion: a manifesto that reveals word by word, discipline cards that stack as you scroll, and a project gallery that settles into place without a single layout shift.
Motion spec · live replica
Concrete is not just a building material. It is a canvas for modern design.
FIG. 4.2A replica of the site’s manifesto reveal, rebuilt here in a few lines of CSS.
§ 03Systems
Behind the products sits infrastructure I built and run, including an AI server we own, wired privately into production. It keeps pages fast, AI costs predictable and sensitive data on our own hardware.
Production · cloud
Data plane
In-house · owned hardware
FIG. 5.1How a request travels. Cache hits return from memory; heavy work is queued; workers reach the in-house GPU over a private Tailscale mesh, so the model server is never exposed to the internet.
Open-weight language models served from a GPU server we own and run.
Why it matters
Inference is a fixed cost rather than a per-token bill, and sensitive text never leaves our hardware.
A private WireGuard network joining the in-house server to production.
Why it matters
No ports open to the internet, yet production reaches the GPU as if it sat on the same network.
A durable queue for scraping, embeddings, model calls and email.
Why it matters
Slow work never holds up a page, and queued jobs survive restarts instead of disappearing.
An in-memory layer in front of the database and the models.
Why it matters
Repeat requests are answered from memory, keeping responses quick and server load flat during spikes.
§ 04Case study · AI platform
For an AI research platform (client name withheld), I reworked three things: how it gathers information, which model answers which question, and how much of the work happens at once.
6.1
Sources are fetched concurrently instead of one after another, cleaned once, and de-duplicated before anything reaches the index, so the assistant reads more pages in less time and never pays to process the same page twice.
Sources
sites, documents, feeds
Fetchconcurrent
Clean
boilerplate stripped
De-duplicate
seen pages skipped
Index
ready to retrieve
FIG. 6.1The ingestion pipeline. Fetching fans out across workers; everything after it runs once per unique page.
6.2
Every request is routed by who is asking and what they are asking. Quick look-ups go to a fast, inexpensive model; real analysis goes to the most capable one. Answers stay sharp where it matters and the AI bill stops growing with traffic.
“When is the filing deadline?”
Free plan
“Summarise this policy for me”
Pro plan
“Compare these three agreements”
Business plan
Router
plan ×
question
Fast model
Short factual look-ups
Balanced model
Summaries and everyday answers
Deep model
Multi-document reasoning
FIG. 6.2Model routing by plan and question type. Example questions shown; the router runs on every request.
6.3
Searching, scraping and retrieval do not depend on each other, so they now run side by side and the answer streams as soon as the model has what it needs. The wait is set by the slowest step, not the sum of all of them.
Before · one step at a time
After · independent steps together
FIG. 6.3Request timeline before and after. Proportions are illustrative; the shape of the change is the point.
§ 05AWS credential
Certified by Amazon Web Services in May 2026 to design secure, resilient, high-performing and cost-optimised systems on AWS. The same principles run through every system on this page.
Core services the certification covers
What it covers, and where it shows up above
01
Private networking over Tailscale, least-privilege access, and Margin’s vault encrypted in the browser.
02
Queues between services, so one slow or failing dependency does not take the product down with it.
03
Caching layers, server rendering and image discipline: 96–100 Lighthouse performance on client sites.
04
Owned GPUs for inference and models matched to the question instead of one expensive model for everything.
§ 06Working together
Hiring a developer is easy. Trusting one with your business is harder. This is how I work, so there is never a week where you wonder what is happening or what it will cost.
Goals, constraints and what success looks like, agreed before anything is built.
You getA written brief and a list of risks
Architecture, scope and estimate, reviewed together.
You getFixed scope, timeline and cost
Short cycles with working software you can click every week.
You getDaily written updates, weekly demos
Performance, security and monitoring checked before go-live.
You getA signed-off launch checklist
Fixes, improvements and scaling as the business grows.
You getA reply within one business day
Daily update · Checkout rebuild
Gurman Singh → you · Tue, 4 Nov, 6:30 pm
Done today
Next
Blockers
NoneMon
Update
Tue
Today
Wed
Update
Thu
Update
Fri
Demo
Launch checklist
4 / 4 signed offFIG. 7.1An example of the written update you receive at the end of each working day.
Questions never sit unanswered while work waits.
Code, repositories, cloud accounts and data are in your name from day one.
Written estimates, and any change in scope is agreed before work starts.
Least-privilege access, encrypted secrets and a security pass before launch.
Architecture diagrams, runbooks and credentials, handed over properly.
Ongoing support, upgrades and scaling, not a hand-off and goodbye.
§ 07About
I started in 2020 with plain JavaScript, the DOM and CSS, and learned how browsers actually render before reaching for frameworks. That foundation is why the sites above stay fast.
Since then I’ve moved down the stack on purpose: React and Node, server rendering and relational data, then cloud architecture, queues, caching and self-hosted AI. Today clients in India, Australia and Canada trust me with products from first sketch to running infrastructure, and I can explain every layer in between, in plain language.
01
Lighthouse, profiling and real numbers come before opinions about speed.
02
Design, build, infrastructure and launch, handled by the person who understands all of it.
03
Clear diagrams, written decisions and code the next engineer can follow.
The path here
2020
Foundations
JavaScript, the DOM and CSS, learned without frameworks
2021
Full stack
React, Node.js, Express, MongoDB, REST APIs
2022–23
Client ownership
Next.js SSR, TypeScript, PostgreSQL, GSAP
2024
Cloud & scale
AWS, Docker, CI/CD, service boundaries
2026
Certified · AI infra
AWS SAA, in-house models, queues and caching
§ 08Contact
Tell me what you’re planning. I reply within a day with honest next steps: what it takes, how long, and what I’d do first.
What happens next
You send a note
A few lines about the idea or problem is enough.
I reply within a day
With questions, and honest first thoughts.
A short call
We agree goals, constraints and timing.
A written proposal
Scope, timeline and cost, in plain language.