Full-stack engineer. Node, TypeScript and React, and the data and AI systems behind them. London, UK.
Sole engineer on a multi-product clinical platform in London: the APIs, the integrations that keep payments and logistics in step, and AI features that reach no patient without a human reading them first. Before that, the pipelines behind PepsiCo's global supply chain, across 25 million product and customer combinations.
Stated in a written reference from LatentView Analytics: the demand forecasting model I developed, tested and deployed directly impacted over $75 million in sales. The letter is available on request.
Weekly volume through the event-driven pipeline built for PepsiCo's global supply chain at LatentView. Processing time went from 13 hours to 8, a 38 per cent cut.
Measured end to end on the PepsiCo demand planning disaggregation pipeline. Both endpoints are stated because a percentage on its own is not checkable.
Products built end to end at By Dr Vali, from schema design to on-call. Four are publicly reachable: bdvfit.com, bydrvali.com, bdvportal.com and bydrvaliportal.com.
Each one gathers what I built against it, the time it happened to somebody in public, what I published about it, and the material that teaches it.
Nothing about a fluent answer indicates whether it is grounded. The system has no way to say it does not know, so it invents, and the invention arrives in the same tone as the truth.
Ask a clinical question and get the passage the answer came from, or a refusal when the corpus does not support one. Built on a synthetic corpus of invented conditions and doses, with a deterministic extractive fallback so it still answers when the model is unavailable.
make it refuse ↗No AI output reaches a patient without a human reading it first.
A harness testing hallucination, factual accuracy, consistency and safety across several providers, with drift monitoring. MSc research, awarded Distinction.
2024 Moffatt v. Air Canada. A support chatbot invented a bereavement-fare policy that did not exist. The airline argued it was not responsible for its own bot. A tribunal held the airline liable for what its agent said. The bot's confident wrong answer became a binding promise.
FRAME Failure-Recovery Architecture for Multi-step Execution
A row cap, a join that multiplies, a key that does not match. The pipeline succeeds, the dashboard renders, and the number is quietly wrong for weeks.
The request may never have arrived, or it may have completed and the response been lost. Both look identical to the caller, and only one of them is safe to repeat.
Data scripts get a fraction of the review that application code gets, and they are the only code that cannot be rolled back by deploying the previous version.
A green run is evidence only if you know what it ran against. Coverage counts lines executed, not assertions made, and a suite nobody can distrust is a suite nobody checks.
Two clocks, two zones, a schedule declared against local time, a day boundary compared the wrong way. Every one of them is invisible until a date arrives that nobody tested.
A cluster bought for traffic that never arrived is not neutral. It is a thing to operate, secure, upgrade and explain, paid for while the product is still unproven.
Ten questions, then a recommendation where every component carries its reasoning and its real alternatives. At small scale it tells you so, because most systems are over-engineered and a recommendation with no visible alternatives reads as a verdict rather than a choice.
answer ten questionsA live clinical retrieval demo: ask a question, get an answer with the passage it came from, or a refusal when the corpus does not support one. Built entirely on a synthetic corpus of invented conditions, drugs and doses, so nothing here implies real clinical guidance. Runs on a free-tier model with a deterministic extractive fallback, so it answers even when the model is unavailable.
make it refuse ↗The 221 topics behind 48 cards carry 141 cited sources. Every one was requested: 132 resolve, and the nine that do not are explained rather than dropped. Three of my own factual claims turned out to be wrong and are corrected in public, and one defect is left unfixed with the reasoning stated instead of quietly shipped.
read the materialEleven check suites run in the build: 45 assertions against the deployed site, 30 on security, 69 email compatibility rules, plus gates on typography, colour contrast at the sizes actually used, content structure, indexing signals and every route returning real text rather than an empty app shell.
the source ↗Open to software engineering roles, London or remote.
Notes on building systems that survive production: what broke, and what the fix cost. Sent when there is something worth sending, and never for anything else. Read past issues first.
Unsubscribe from any email, in one click. Your address is stored to send this and nothing else, and is never passed on.