§0

Built from scratch, down to the attention weights.

Rahul Malik — AI developer & full-stack engineer, New Delhi.
Co-founder & Technical Lead, Global South Solidarities, a registered Washington State nonprofit.

Abstract

I design and build AI systems and websites from first principles: a one-billion-parameter language model trained from scratch on consumer hardware and free cloud quota,¹ retrieval infrastructure over 1.3 million document chunks,² and a hand-coded, zero-tracker web platform used in four university courses.³ I am taking commissions for custom AI systems, hand-built websites, and data pipelines — see §5. Every figure in this paper computes live, in your browser. Nothing here asks to be taken on faith.

·this page will annotate its own typography

you are reading the inside pages — the front page is typeset by the machine itself →

§1 Selected results

The ledger, before the prose. Each row links to its evidence.

Figure	Claim	Evidence
10⁹ params	Language model designed and trained entirely from scratch — decoder-only, grouped-query attention, RoPE, SwiGLU — on consumer hardware plus free cloud quota. In beta ahead of public deployment.	site →
257,918 articles	Self-curated training corpus from 50+ global news outlets, roughly 15 billion tokens, collected and cleaned with my own pipeline.	site →
1.3 M chunks	Production retrieval-augmented generation pipeline — FAISS, multilingual BGE-M3 — behind a working research assistant.	Plate II →
1 paper	Provenance-Conditioned Attention — source-type gating for epistemic reasoning in transformers. Co-authored with Dr. Bibhushana Poudyal, published on Zenodo; follow-on distillation work in progress.	DOI →
32 / 250	Kaggle, “Harmonizing the Data of your Data” — top 13% of 250 teams, building an LLM-assisted pipeline extracting experimental metadata from scientific publications. One of 10+ international competitions entered.	Plate IV →
3 grants	Two Washington State University awards (sole builder of both funded deliverables) and USD 2,000 in compute from Modal, secured directly, for the 1 B model.	Plate I →
40+ pages	Hand-coded public platform — research publications, course sites, educational games, investigative reports — zero trackers, deployed on Vercel, in classroom use at WSU.	visit →
1 library	nerscope — pip-installable Python NLP library for document comparison and entity extraction, with a full test suite and CI/CD.	PyPI →

To commission work in this ledger's vein — §5.

§2 The marginal model

Portfolios assert; figures compute. The transformer below was trained for this page and runs its forward pass in your browser — hand-written matrix arithmetic in plain JavaScript: no library, no API, no server.⁴

Manuscript line

seed prompts:

Context — red underlines: attention from the next prediction

Probability specimen — candidate next characters, sized by probability

Print density — temperature 0.80

0.2 letterpress1.2 loose ink

Top-k 12

1 strict40 open

JavaScript is off, so the model cannot run — here is a recorded generation from the same weights (prompt “The night ”, temperature 0.8):

“The night of love, the shrink, which is endless, its weight is the poems, and the tears of the flowers are flame.”

Fig. 1. A 600,448-parameter decoder-only transformer — the same family as the author's 10⁹-parameter model (RMSNorm, rotary position embeddings, grouped-query attention, SwiGLU, weight-tied head) at roughly 1/1,600 scale — trained for this page on the selected English works of Rabindranath Tagore,⁵ quantised to int8, and executed by ~400 lines of hand-written JavaScript. The point is the machinery, not the prose.

Checkpoint — drag to watch it learn

step 02001,0006,000

Fig. 2. The run itself. Validation loss 117.98 → 1.35 nats/char. Each checkpoint is the real weight file from that step; sampling is live inference, so you can hear the model pass from noise, through word-shapes, to sentences.

4Open DevTools → Network. The weights arrive as one 624 KB binary; after that, every token is local arithmetic. View source: js/model-worker.js — the forward pass is readable.

5881,952 characters of public-domain prose and verse: Gitanjali, Sadhana, The Home and the World, Stray Birds, The Gardener, Fruit-Gathering, The Crescent Moon. The model babbles like a poet because a poet is all it has read.

Components, named

RoPE — rotary position embeddings; positions as rotations in attention space.

GQA — 4 query heads share 2 key/value heads; the memory-saving trick used in production LLMs.

SwiGLU — gated feed-forward, the modern default.

KV-cache — past keys/values are kept, so each new token costs one position, not the whole context.

On conditioning attention itself: Provenance-Conditioned Attention (Zenodo).

§3 Plates — selected commissions & builds

PLATE I Planetary Justice Map

A university needed its research on transnational solidarity to be teachable — not a PDF, but a place students could walk around in. Funded by a Washington State University research grant,⁶ I designed and built the whole deliverable solo: an interactive 3-D globe connecting ten revolutionary figures and movements across five continents with animated solidarity arcs. Every node opens an archival dossier with original scholarly writing, thematic tags, and a navigable web of connections; the whole thing is layered so future funded projects extend it without a rebuild.

It shipped with thematic filters, a classroom index panel, a mobile layout, and reduced-motion support — and it is in use in university teaching now.

What this proves for you: complex, interactive data products — scoped, built, and delivered by one person, on a grant deadline. Commission a build like this — §5 →

Exhibit. A working miniature, drawn for this page — orthographic projection, coastlines, and solidarity arcs in ~200 lines of hand-written canvas code, no mapping library. Drag it.

PLATE II Retrieval at production scale

You have ten thousand documents and no way to ask them a question. That is the problem retrieval-augmented generation solves, and I run one in production: 1.3 million document chunks embedded with multilingual BGE-M3, indexed in FAISS, feeding a research assistant that cites its sources — so every answer can be checked, not merely believed.

What this proves for you: your archive — contracts, tickets, research, catalogues — can answer questions in plain language, with receipts. Commission a build like this — §5 →

Exhibit. The pipeline, as built.

PLATE III gssolidarities.org — the platform

Forty-plus pages, every one hand-coded: research publications, course sites for four university classes, educational games, an investigative journalism section, long-form multimedia reports, and a contributions archive publishing work by 10+ graduate students. One consistent design system. No page builders, no themes, and — as a matter of principle — no analytics, no trackers, no cookies. It deploys as static files on Vercel and it is fast because there is nothing in it that does not need to be there.

pages	40+, hand-coded
trackers	0 — verifiable in your DevTools
classrooms	4 WSU courses, each term
stack	HTML · CSS · vanilla JS · Vercel

What this proves for you: an entire public platform, designed, built, and operated by one engineer — no agency overhead, no handoffs, no lock-in. Commission a build like this — §5 →

PLATE IV Short notes

nerscope. A pip-installable Python NLP library for document comparison and entity extraction — full test suite, GitHub Actions CI/CD, on PyPI.
Kaggle briefs. 32nd of 250 teams (top 13%) extracting experimental metadata from scientific papers with an LLM-assisted pipeline; 10+ competitions across bioacoustics (BirdCLEF), machine translation of Old Assyrian cuneiform, legal information retrieval, and LLM reasoning benchmarks — custom models end to end.
Programmatic publishing. A print-ready investigative zine produced by a ReportLab PDF pipeline — the same machinery that can generate your reports, catalogues, or certificates from data.
3-D animation. Two documentary shorts in production in Blender — on the Rohingya genocide and Facebook's algorithmic role, and on cobalt mining in the DR Congo.

What this proves for you: range — the person who trains the model can also package the library, cut the film, and typeset the report. Commission work like this — §5 →

§4 Method

From scratch is a discipline, not a flex. Knowing every layer means nothing in the stack is magic, and nothing that breaks is unfixable.⁷
Privacy is a default, not a tier. No trackers, no analytics, no third-party calls unless the work demands them — and then, documented.⁸
Evidence over adjectives. Claims carry links. Deliverables ship with the means to verify them.⁹
Small budgets are a design constraint, not an excuse. The billion-parameter model was trained on consumer hardware and free quota. Constraint is where the craft is.
Ship things people can inspect. Hand-written code, readable source, documented handover. You own what you paid for.¹⁰

§5 Commissions — services & terms

You hire the person who did everything in §1 — no subcontractors, no templates. Three standing offers:

A · Custom AI systems

You have documents, data, or a workflow that needs a mind attached to it.

Work: retrieval pipelines (RAG) over your archive · model evaluation & fine-tuning · LLM-powered features built into your product · Claude API integrations done properly
You receive: scoping memo → working system → evaluation report → documented handover
Evidence: Fig. 1, Plate II

Enquire →

B · Websites built from scratch

No page builders, no bloated themes — a site written by hand, one you own outright.

Work: portfolio & studio sites · publications & archives · course platforms · interactive data-driven pages
Terms: performance and privacy budgets written into the agreement — measured page-speed targets (Google Lighthouse) and a zero-tracker guarantee as contract terms
Evidence: Plate III, the Studio, and the page you are reading

Enquire →

C · Automation & data pipelines

The repetitive thing your team does every week can be a script that runs every night.

Work: scraping & data curation (built a 257,918-article corpus) · document processing & NLP tooling · programmatic PDF/report generation · publishing pipelines
You receive: a running pipeline, scheduled and monitored, with source and docs
Evidence: Plate IV

Enquire →

Pricing is by written estimate after a free scoping call — a fixed quote within 48 hours, no meter running.

Fit check

right fit — founders and teams who want to own their stack, not rent it
right fit — creators & researchers who need fast, custom, careful work
wrong fit — “just install a template” — you would be overpaying me

Research collaborations — co-authorship, teaching tools, grant-funded builds — are considered separately; write with a one-paragraph abstract.

§6 Correspondence

To commission work or propose a collaboration, write:

rahul.malik11011983@gmail.com

Or let the page draft the letter — tick what applies:

Open in your mail app

github	github.com/RahulMalik22 · github.com/gssolidarities
organisation	gssolidarities.org
research	Provenance-Conditioned Attention — Zenodo
library	nerscope on PyPI

correspondence answered within 24 hours