Ten years shipping 0→1 across five markets. I build products where constraints live — attribution pipelines, inference economics, measurement systems, and delivery optimization — and turn them into multi-sided platforms.
One platform case study, then things built in the open — code and live demos on GitHub.
Why decode workloads run at roughly 0.2% of a chip's theoretical potential: the bandwidth gap between HBM and the MXU. Ships an interactive calculator and a Python arithmetic-intensity estimator. Part two works through quantization, batching, speculative decoding and KV-cache reduction.
A GTM operations console for AI product portfolios: track the freshness of the public documentation field teams depend on, monitor competitive pricing moves across AI labs and vendors, and assemble evidence-based action memos.
Try it live →Server-side bot detection pipeline: Cloud Armor + reCAPTCHA → BigQuery → Vertex AI clustering → $0 value-based-bidding conversions for bots. Fixes bot contamination at the infrastructure layer so Smart Bidding learns to avoid bots on its own, instead of patching it downstream in reporting.
Read the architecture →Apple's App Tracking Transparency broke the link between an app advertisement impression and an in-app conversion, reducing conversion visibility for advertisers.
An end-to-end platform: deeplinking for deeper shopping journeys, device-ID attribution, bidding integration, and in-app conversion modelling stitched into one pipeline. Shipped the iOS ATT recovery path, then graduated the whole ecosystem onto YouTube Connected TV.
I ran experiments to measure the conversion lift for advertisers. The ATT fix drove eight-figure incremental ARR; the platform now underpins multi-billion-dollar annual advertiser spend.
An agent for award-seat availability and credit-card transfer-partner optimization across airlines. Gemini API behind a serverless Vercel proxy, so the key stays server-side and never ships in the client bundle.
Try it live →Lead cross-product enterprise segmentation and go-to-market for Search, YouTube and ROI measurement. Built an “Enterprise Sentiment Score” from scratch on a RAG + RLHF pipeline — now used by 3,000+ internal users across a nine-figure portfolio. Scaling adoption of Google's generative AI marketing products (FeedGen, VideoGen) through agencies and third parties.
Owned Web to App Connect, the attribution infrastructure underpinning multi-billion-dollar annual YouTube advertiser spend: deeplinking, device-ID attribution, bidding integration, in-app conversion modelling. Shipped an iOS ATT recovery path generating eight-figure incremental ARR and graduated the ecosystem onto YouTube Connected TV. Designed and ran the experiments used to measure advertiser conversion lift.
Full-time MBA between operating roles.
Built NowNow, a hyperlocal delivery marketplace, from zero to 3,000 brand partners and 200,000 customers. Launched a VIP loyalty program personalized with Google Recommendations AI. Worked directly with the CEO and CSO on the materials and models behind a multi-billion-dollar capital raise.
Omni-channel marketplace connecting smallholder farmers directly to market. Built the acquisition strategy, channel partnerships and growth operations from scratch.
Fast-track promotion, one of seven globally from a cohort of ~110 analysts. Scoped 60 digital products — AI, IoT, blockchain, AR/VR, robotics — for a $500B smart city, with roadmaps, reference architecture and commercialization plans adopted by a Middle East government. Co-founded BCG's blockchain sub-practice.
Research on the India–US Defence Technology and Trade Initiative, supporting the policy analysis behind the ten-year Defence Framework Agreement between the two governments.
Framework agreement ↗Market-access platform for artisan communities. Raised $10K and grew sales 50% through NGO channel partnerships.
The roofline model applied to LLM inference — arithmetic intensity, the memory-bound zone, and six levers for un-stranding a chip that's sitting idle waiting on HBM.
An interactive amortized-roofline calculator: set the chip, model shape and batch size, and see where throughput and cost-per-token actually land.
Why filtering bots in the reporting layer teaches your bidding model nothing, and what a server-side detection pipeline into Vertex AI looks like instead.
Peer-reviewed IEEE work on ML-optimized filter design using Firefly and Cuckoo-Search algorithms. Where the engineering foundation comes from.
↗
Press
Noon VIP Loyalty Program LaunchAl Bawaba · Saudi Arabia · 2021
↗
Policy
India–US Defence Technology & Trade InitiativeTen-year Defence Framework Agreement · researched at India's Ministry of External Affairs · 2016
↗
Research
IEEE — IIR Digital Fractional-Order Differentiator & DC BlockerML-optimized design using Firefly & Cuckoo-Search algorithms · 2018
↗
Hardware
Kuhn's Poker on an Intel 8085 MicroprocessorGame-theoretic solver implemented on a circuit board · NSIT Delhi · 2017
↗
I started in electronics engineering, which helps me build tangible products that solve for constraints — memory bandwidth, signal integrity, marketplace supply.
I love to build and operate in places where the playbook needs to be written from scratch: a delivery marketplace across three Gulf markets, a blockchain practice inside BCG, an attribution platform at Google. I'm most useful in building in ambiguous spaces, in connecting teams towards synergies, and influencing through narratives.
Ten years across the US, India, UAE, LATAM and Canada.