Maneki-neko dangle selected.
Brand systems. Research interfaces. GTM. Films. Essays. Drawings in fine pen. Open-source contributions. And no working sense of when to stop.
These projects begin where the pitch deck politely leaves the room: the product has to work, and the person using it has to know what to do next.
JunoAgent system · Marketing
Juno is an AI-native content marketing operating system, running continuously rather than campaign to campaign.
A dreaming engine, a team of specialist content agents, and a shared data foundation of company memory and GEO/AEO intelligence — working together to discover, research, produce and improve the content that moves the business.
BardAgent system · Field marketing
Bard is a lead-enrichment and relationship-intelligence platform that turns target accounts into the right people, relevant context, and credible ways to reach them.
Coverage measured against a single-source enrichment workflow; accuracy across evaluated names, companies and roles, with every detail traceable to its source.
bobDesign system · Payments
bob is a checkout builder for UPI. Most checkout builders start with components; bob starts with the journey a merchant is trying to make, showing UPI Intent, Collect and Autopay before anyone commits to the build.
Three UPI journeys have been designed: Intent, Collect and Autopay. One identity system carries all three from journey selection into the builder.
DexterRetrieval system · Research
Dexter is a second brain for a payments company. Payments research tends to end in six tabs, three PDFs and one heroic colleague with the correct link. Dexter keeps the answer, its source and the next question in one branching thread, so a team can go deeper without losing the way back.
Measured over 134,758 retrievable passages drawn from 14,769 documents and 115 distinct sources. The public experience is deployed with synthetic and public-source material.
Data story · Podcast archive
I turned more than four million words of public Lenny's Podcast transcripts into a data story, an eight-question product quiz and collectible guest cards. It won the lennysdata.com hackathon. Pokémon TCG was referenced heavily. Almost certainly to an unhealthy degree.
Beautiful, fun, creative, and well executed.
DreamBench-SWE asks whether software agents can remember the right thing at the right time — and turn it into working code. Most memory benchmarks stop at recall: did the agent remember the fact? This one keeps going until that fact has survived several sessions, plausible distractions and an executable code test.
Four conditions over 180 executable tasks each: no external memory passed 21, verbatim event memory 82, a typed-plus-raw reference probe 83, and one pinned hosted configuration 97. Every memory-bearing condition beat no-memory after Holm correction.
Skill library · Product marketing
PMM OS is a public library of twenty-eight product-marketing skills with the method left visible. Skill libraries are very good at names; then the useful part disappears behind a prompt box and a large ‘trust me.’ This one shows what to bring, how the work runs, where it can fail and what comes out.
28 original skills across seven live categories. The measured page records 72,413 tool invocations and 2,679 MCP calls across 30 days.
Complex products do not become clear because somebody put a gradient behind the acronym. This work gives each one a public idea people can understand and remember.
Global brand · Go-to-market · Positioning
Juspay had become many products in many markets, but the brand still behaved like a collection of separate parts. I led the visual rebrand and website direction to give the company one recognisable system, with enough room for different payment worlds to stay different.
A new identity and website system across product stories, motion and regional expression — made with a wider team.
Brand · Go-to-market · Positioning
Payment infrastructure tends to disappear until somebody needs to change a route and discovers a wall of machinery. I led HyperSwitch's brand strategy and creative direction around a clearer promise: make control visible.
Brand strategy, creative direction, website direction and motion — built with visual, motion and product collaborators.
Brand · Go-to-market · Positioning
Brazil already knew PIX. The job was not to arrive with a lanyard and explain it back to the country. I shaped the brand and GTM thinking around what Juspay could add: clearer product stories for PIX Biométrico and PIX Automático.
Market-entry narrative, product explanation and campaign worlds for PIX Biométrico and PIX Automático.
Product narrative · Brand system · Go-to-market · Product
Breeze Automatic is an AI sidekick for founders: it turns the questions hiding across dashboards — revenue, traffic, orders, ads, payments — into a clearer next move. Most business dashboards cast the founder as supporting talent in a software demo. Breeze Automatic keeps the founder in charge, turning business signals into a useful next move without asking them to become a full-time dashboard operator.
Work shown: positioning, product narrative, identity, motion and launch expression.
Cross-border payments · GTM film
Cross-border payment stories usually open with routes, rails and settlement, then remember halfway through that somebody was trying to finish a song. I led this film in the opposite order: meditation, music and play establish what payment is protecting before the machinery enters.
Creative direction and narrative, visual & motion design made with Figma Weave.
XyneBrand · Positioning · Go-to-market
Xyne is Juspay's enterprise AI platform. Enterprise AI has a habit of arriving one tool at a time, like house guests who each bring a key and then ask who controls the thermostat. For Xyne, we made control the organising idea: one governed plane beneath the assistants, models and tools, drawn so governance is something you can follow.
Creative direction, strategy, positioning and go-to-market.
Software · Open source
I had fallen for open source rather late and wanted to give something useful back to projects I had used, admired and learned from.
Merged work, by project. Star counts checked 22 August 2026.
Block / Buzz · 29.3k · #5143: fix(desktop): drop unhandled rejection from throwing window.Notification, #5142: fix(desktop): fence localStorage SecurityError from killing the React tree, #4374: fix(desktop): stop clipping focus ring on channel intro action cards (#2392), #4304: fix(desktop): rename generic attachment action from ‘Attach image’ to ‘Attach file’ (#2381)
Goose · 53.2k · #10364: fix(mcp): keep stdio extensions alive across worker exits
vLLM / Semantic Router · 5.2k · #2841: [Router] Poll for Milvus cache visibility in integration test, #2842: [Router] Honor replay Milvus consistency level, #2812: docs: automate API reference generation from route catalog, #2510: [CLI] Contain runtime state files under owned directory
Vercel / SWR · 32.5k · #4310: fix: clean up completed subscription state
LangChain · 144.7k · #39338: fix(openrouter): preserve cost metadata in usage chunks
The honest answer is I don't know, and I start anyway.
The writing is trying to prove me wrong. The drawing is not trying at all.
2608.20664