I build the systems that are still standing at 3am.
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.
experience
Builds public and admin REST APIs (Node.js/Express, OAuth, RBAC, rate limiting to 5,000 req/sec/IP) and webhook-driven integrations across Stripe, Magento, and Intercom with idempotency and retries. Led a full Magento rebuild and shipped 13+ products end to end, from schema design to on-call.
Architected event-driven distributed pipelines processing 25M+ weekly records with PySpark on Azure Databricks for PepsiCo's global supply chain; cut processing time 38% (13h → 8h). ML segmentation for Hilton delivered +26% retention and $10M annual uplift. Clients: PepsiCo, Hilton, Unilever.
selected work
API Platform & Integration Hub
Production REST API platform: OAuth, role-based access, and rate limiting at 5,000 req/sec/IP, integrating Stripe, Magento, and Intercom via webhook workflows with idempotency, retries, and failure handling at live scale.
DPD, Demand Planning Disaggregation
Event-driven distributed pipeline processing 25M+ weekly records for PepsiCo's global supply chain. Cut processing time 38% (13h → 8h); contributed to systems with approximately $60M in annual revenue impact.
Customer Retention Analytics
Segmentation and behavioural pattern models over 100+ features for Hilton. Delivered a 26% retention improvement and $10M annual revenue uplift.
CPD-accredited Level 7 platform for doctors and dentists: 25+ screens, 20+ API modules, video lessons, quizzes, Stripe payments, auto-generated certificates, and an 18-page admin panel. Built solo in six months.
Full Magento platform rebuild supporting 120+ services and high-volume traffic, with Docker environments and CI/CD pipelines for reliable, fast deployments.
A 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.
LLM Response Evaluation Framework
Automated harness testing hallucination, factual accuracy, consistency, and safety across multiple LLM providers, with a model-drift monitoring dashboard. MSc research, awarded Distinction.
writing
all writing, and the feed →The Unanswerable Test Suite: How to Test AI Systems That Sound Right When They Are Wrong
Argues that testing an AI system needs a suite built from questions it should refuse, not just ones it should answer, opening with a consulting firm's AI-generated report caught citing sources that never existed. Proposes tracking refusal rates directly and anchoring factual claims to source documents with deterministic checks, not another model's judgement.
The Quiet Failure Mode of AI Features: When Your Cached Outputs Go Stale
Traces a semantic-search bug to a bulk update that bypassed the save hook, leaving cached embeddings stale with no visible error. Lays out a three-layer defence: mark on change, recompute in the background, and run a reconciliation sweep that catches whatever the first two miss.
You Might Not Need a Hosted Embedding API
Opens with the 2025 Shai-Hulud npm worm harvesting credentials straight out of build environments, to question why small corpora default to a hosted embedding API. For a few thousand documents, argues for a local model behind an abstracted scoring layer: no per-query cost, a smaller attack surface, and room to graduate to a hosted service if scale ever demands it.
The JVS Audit: Why Your Cross-Platform Dashboards Are Quietly Lying to You
A tool-agnostic audit for cross-platform dashboards, opening with the 2020 case where nearly 16,000 COVID results were lost to a spreadsheet row limit. JVS is Join, Validate, Standardise: catch silent row multiplication, measure key match rates, and hold one authoritative definition per metric.
Design your AI agents around how they fail, not what they can do
Argues agentic systems should be designed from their failure modes, opening with the Moffatt v. Air Canada chatbot case. Introduces FRAME: failure classification, recovery logic, awareness boundaries, monitoring hooks, and an escalation protocol.
Treat every data migration like it will go wrong
Makes the case for applying the same rigour to data scripts as to application code, opening with the 2017 GitLab data-loss incident. Advocates dry-run-by-default, idempotent markers, paired verification scripts, and a migration log.
Time is a feature: One decision that removes a whole class of bugs
Treats time as a design decision, opening with the 2012 leap-second outage. Store timestamps in UTC, choose one human-facing timezone, declare schedules against a fixed zone, compare against start-of-day, and freeze the clock in tests.
podcast
Trustworthy AI in Healthcare and Longevity ↗
A guest conversation on what it actually takes to earn trust for AI in healthcare and longevity research. Covers where a confident-sounding model still needs a deterministic check behind it, and how that changes what you test before shipping.
Building Safe, Testable AI in Healthcare & Longevity ↗
A follow-on conversation for a testing-focused audience, covering the same healthcare AI territory from the QA side: what 'safe' has to mean before a testable claim about a system can be made at all.
speaking & recognition
Designing APIs and Integrations That Don't Fall Apart at Scale
Judging panel, and presented two awards on stage at the 2026 ceremony
education
GPA 8.5/10. Dissertation (Distinction): anomaly detection across S&P 50 equities, comparing LSTM and GRU networks and looking for irregular market behaviour rather than forecasting price. Prices pulled from the Yahoo Finance API into S3, retrieval automated on EC2, deviations flagged with ADTK, and the results published to a QuickSight dashboard. The LSTM trained to 2020 found anomalies most reliably. Coursework covered spatial data analysis, advanced decision making, quantitative data analysis, and a module on how governments should regulate AI, which is closer to the day job than it sounds. Access and Participation Board member.
GPA 9.26/10, all 82 credits earned. The thesis was the largest single piece of the degree at 16 credits and graded S, the highest VIT awards: object detection by transfer learning on CIFAR-10, reaching 96% accuracy, published as an IEEE paper. Same grade in database technologies, applied statistical methods, distributed operating systems and object-oriented software engineering, which is most of what the pipeline work later stood on. Coursework ran through machine learning, data mining, big data analytics and advanced software testing, which stuck. Merit Scholarship recipient.
1654/1900 overall, at Dr. D. Y. Patil Arts, Commerce & Science College, Pune. Compiler construction, operating systems, computer networks, system programming and computer graphics, alongside 8051 architecture and interfacing, which is the part that made the hardware work. Built a line-following robot at a robotics workshop run in collaboration with IIT Delhi: infrared sensors reading the contrast under the chassis, a microcontroller correcting the steering continuously, and a control loop that had to be tuned rather than calculated.
newsletter
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.
Unsubscribe from any email, in one click. Your address is stored to send this and nothing else, and is never passed on.