Skip to content

(1)Work

Things I've built, with the decisions written down: what, why, and what it cost.

A browser frame labelled fiwoodart.co.in/collections holding a wireframe of the catalogue: a round logo, five navigation lines and an orange Enquiry button with a count of three, a column of filter checkboxes with one ticked, and four cards marked F1 to F4, each holding the same winged silhouette. Beneath, the caption The catalogue as a wireframe, no cart, one enquiry list, sent on WhatsApp, the line Browse, then ask, and the note 203 pieces.
[01]2026

A family wood-art business, from a PDF catalogue to a website with WhatsApp in place of a cart

The first website for FI Woodart, my family's business, which handcrafts layered 3D Zoroastrian wood art in Pune. Until October 2026 the whole range lived in a 51-page PDF catalogue. I designed and built a site that turns it into 203 pieces a visitor can browse by collection, wood shade, format and price, then sends whatever they pick to the makers as one ready-written WhatsApp message instead of through a checkout. It went live at fiwoodart.co.in on 15 October, a day after I started, and I have kept it up to date since.

203 pieces online, and every enquiry arrives as one WhatsApp message

WebsiteCatalogueStatic site
Three rows of blocks. Eight outlined boxes named for the prompt modules, from identity to examples. Twelve solid grey blocks for the knowledge-base files, all of them. Then a dashed box for the conversation history and one solid accent block for the new message. The first two rows are labelled as the cached prefix, byte-identical on every call
[02]2026

A WhatsApp sales bot with no RAG: the whole knowledge base rides in a cached prompt

A WhatsApp sales and support bot for Divo, an Indian car dashcam brand, built as a freelance project and live on the brand's main WhatsApp number since 30 September 2026. It answers product, installation, warranty and policy questions only from a 12-file Markdown knowledge base, recommends one camera for the customer's car, and sends anyone who wants a person to the support line. There is no vector store: the whole knowledge base sits in a cached prompt prefix, so prompt size is the bill, and on the live service a cached reply costs about $0.0004.

Live since 30 September, about $0.0004 per reply, measured

AI agentConversational AILLM cost engineering
A browser frame labelled Nextoar, preview, holding a wireframe of the home page: a header with a logo, six navigation lines and an orange pill button reading Request a Demo, then a hero made of five slanted grey photograph panels with two heading bars across them, and beneath it a line of copy, an email field and the same orange button. Below the frame, the caption The home page as a wireframe, plain HTML with a little script on top, the line Static, on purpose, and the note 25 KB of script.
[03]2026

A startup's website, built twice: first in Next.js, then as plain static pages

The new website for Nextoar, the industrial AI startup in Bangalore where I work as an AI engineer intern. I designed and built it to replace a WordPress site that the company had outgrown: five products where there had been none on the menu, a demo request within reach on every page, and nothing to patch. I built it in Next.js, then moved the whole thing to Eleventy, because every page is the same for every visitor and a framework was weight it did not need. It is finished and on a preview address, waiting for the switch to nextoar.com.

Every interaction on the site runs on 25 KB of JavaScript

WebsiteStatic siteFront end
Ten bars for ten years of price-to-earnings multiples. The first four are drawn in grey at a quarter of their accent outline and the next three at half, with dashed lines marking a bonus issue and a stock split. The last three bars are solid, where the two calculations agree
[04]2026

Every stock came back 'wait'. The prices were split-adjusted and the earnings were not

StockLens is a local research tool I built for my own investing in Indian listed stocks. A Screener.in export and a company's transcripts go in, and a cited report ending in a stance comes out. The arithmetic is plain Python, a 10-node LLM pipeline writes the narrative, and code rejects any claim whose source id is not in the evidence. One user, 28 stored reports across 18 stocks as of 2 September.

One bank's median P/E went from 8.7x to 20.6x, and its entry level from 33% below the price to 5.5%

AI agentLLM pipelineFintech
A browser frame labelled shipsurveyorsfujairah.com holding a wireframe of the home page: a header with a logo, six navigation lines and an orange Request Survey button, then a grey hero block with two heading bars, a line of copy and the same orange button beside an outlined one. Beneath, the caption The home page as a wireframe, one orange button, the same one on every page, the line Found, and fast, and the note On screen in 0.87 s.
[05]2026

A marine surveyor's website, rebuilt to be found and to be fast

The public website of Constellation Marine Services' Fujairah branch, designed and built during my apprenticeship in Dubai and live since June 2026. It replaced an old site that was slow, dated and hard to maintain with a Next.js site built around one job: be found by the people who need a surveyor at the Port of Fujairah, and get them to a survey request in one tap. Six service groups, an experts page, a gallery and 28 articles carried over with their old addresses redirected.

Home page on screen in 0.87 s, nothing moves as it loads

WebsiteSEOPerformance
Rows of dark bars like lines of rule text, bracketed into five chunks labelled with clause numbers, with the chunk for FINRA 2210(d)(2) picked out in accent
[06]2026

Four chunking strategies left recall near 0.2. Reading the rulebooks took it to 0.89

A LangGraph pipeline that takes a finance influencer's YouTube video and returns timestamped, rule-cited compliance findings against 52 pages of SEC, FTC and FINRA text. On 18 hand-written cases the retrieval and citation layer scored 0.944 context recall and 77.8% exact citation match. The full video-to-verdict pipeline was never scored.

Context recall 0.21 to 0.94 on the same 52 pages and 18 test cases

RAGAI agentCompliance
Four boxed tokens reading Every, effort, moves, you feed an arrow into a row of twelve identical tall blocks, each holding a dark attention layer over a grey feed-forward layer and numbered 01 to 12, with one accent-coloured box for the next token coming out on the right
[07]2025

GPT-2 small, written from scratch before I let myself use one

Five weeks in the summer of 2025 spent writing GPT-2 small in PyTorch, one piece at a time: a tokenizer, attention in four versions, the transformer block, a training loop, then OpenAI's released weights loaded into my own architecture, a spam classifier fine-tuned to 96.0% test accuracy, and GPT-2 medium instruction-tuned and marked by a Llama 3.1 judge at 63.8 out of 100.

OpenAI's GPT-2 weights loaded into the architecture I wrote, and it spoke English

LLMDeep learningFine-tuning
A grid of small dark cells like an expression heatmap, with one gene row picked out in accent
[08]2024

Classifying breast tumours from 19,644 genes, and finding the few that matter

M.Sc. dissertation at the University of Liverpool: molecular subtype, receptor group and histological type of 3,207 breast tumours predicted from bulk RNA-seq alone, with differential expression and permutation importance pointing at the genes doing the work.

93.8% accuracy on histological type, from expression alone

ClassificationBioinformaticsResearch
[02] Contact

Want to talk? Say hello.

A project, a question, or just to compare notes. I read everything and write back within a couple of days.

Write to meEmail