Work

I've worked in tech since 2015, almost always at early-stage startups. I joined Harness, Giga, Coram AI, Dgraph and GM Labs early, building their products and the systems under them from zero to one. I also founded Superstream, and before all of them Plato, while still at college.

Between roles I work independently, consulting and doing my own research. The projects I'm proudest of are under notable personal projects below, and the rest are on projects.

• Independent consulting and research

  Jun 2026 – Present, Aug 2024 – Sep 2025

  I work independently, building AI agents for clients and doing my own research. My projects cover evals, multimodal embeddings, pretraining, voice agents and durable execution, mostly in AI, and I've recently started on robotics with an SO-101 leader-follower arm kit.

  The projects are under notable personal projects below.

• Coram AI

  Founding software engineer, Oct 2025 – May 2026

  Coram AI builds AI-powered physical security and operations software, turning ordinary video cameras and sensors into intelligent systems that help organisations improve safety, security and operational efficiency.

  I built Coram Assistant, where customers ask anything about their security infrastructure and run long-horizon investigations into what is happening in their buildings. I also worked on platform features like real-time alerts and our background tasks framework.

• Giga

  Founding software engineer, Jan 2025 – Mar 2025

  Giga is an enterprise-focused AI platform that builds voice and chat AI agents to automate and scale customer support and operations for large companies. Customers include DoorDash and Cloudbeds.

  I built the voice AI infrastructure at Giga, which was until then a text- and email-based customer support platform. The work involved solving problems such as RAG latency on the knowledge base, user interruptions and accents, adding fallbacks for our AI providers and making every part of the voice experience configurable.

• GM Labs

  Founding software engineer, Apr 2022 – Apr 2024

  GM Labs built four products in two years:

  • Perch, an app for reading long-form blogs and essays
  • Startup Archive, curated startup advice from the world's best founders
  • A second Perch, an AMA tool for creators
  • gm.xyz, a place to host your community

  I was a founding engineer responsible for ideating and implementing products from scratch. I led backend development, infrastructure setup and maintenance, AI-based article content scraping and indexing, an article recommendation engine, web3 token gating and web and mobile UI.

• Superstream

  Founder, Sep 2021 – Mar 2022

  Superstream is a protocol and a collection of SDKs for real-time money streaming on Solana. It allows anyone to continuously send money to anyone else at any interval down to a second, and companies used it for payroll and subscription payments.

  The motivation was the earned wage access problem. Low-income workers sometimes face a cash crunch at the end of the month when there is a sudden expense, and have to resort to predatory loans while they wait for a payday that rarely lines up with their needs. Superstream lets workers get paid hourly, daily or at any other cadence, and lets people pay for services in smaller increments instead of monthly or yearly.

  I built the streaming payment system: I designed and implemented the protocol, deployed it on Solana and wrote its SDKs. The project is open source and documented at superstream.finance.

• Harness

  Founding software engineer, Apr 2019 – Aug 2021

  Harness is a modern software delivery platform helping enterprises ship code easily, quickly and resiliently, now valued at $5.5 billion. I joined as one of its earliest engineers, and while I was there it grew revenue 50x and went from one product to six.

  I was part of the Continuous Delivery and Pipeline teams, building the foundations every other team (CD, CI, Feature Flags and Cloud Cost Management) used to build workflow products. My main work was the design and development of the workflow execution engine under all of these products, which runs durable workflows that last weeks. I also led the move from a monolithic Java app to polyglot microservices as the product offering grew.

• AlphaGrep Securities

  Quant researcher and trading analyst, Jul 2018 – Mar 2019

  AlphaGrep is a quantitative trading and investment firm.

  I researched and developed high-frequency quantitative trading strategies for commodity markets in China, which became my team's most profitable market.

• Dgraph

  Founding software engineer, Nov 2017 – Jun 2018

  Dgraph is an open-source distributed graph database with over 20,000 GitHub stars. It is built to be distributed and scalable.

  I was a founding engineer at Dgraph. I launched Ratel, which is Dgraph's data visualisation and cluster management interface. I created tooling for the ecosystem. I built JavaScript (browser and Node.js) and Python clients for Dgraph and contributed to the other clients. I'm still the top contributor to Ratel and the JavaScript and Python clients. Those clients, with the JavaScript HTTP client for browsers, get about 100,000 downloads a month together. I also laid the groundwork for the Dgraph-as-a-service platform built on top of Kubernetes.

• Plato

  Founder, Jul 2015 – Sep 2017

  Plato is a chat app for one-on-one career mentoring.

  I founded the company while still at college, after a pilot during placement season, and worked on the product and tech. It was funded by Commence Mint and selected for Y Combinator Startup School. We came second in Microsoft Imagine Cup, India (2015) and were incubated at IIT Kanpur.

Notable personal projects

• Robotics

  2026 – Present

  I recently started hacking on robotics, beginning with the SO-101 leader-follower arm kit: a leader arm I move by hand and a follower arm that copies it.

• design2code-rl

  2026

  design2code-rl is a synthetic-data and RL-environment pipeline. It generates multi-page websites from scratch and grades how well a coding agent rebuilds them from screenshots, grading the rendered pages, text, code quality and consistency. It runs on Harbor and Modal over 50 generated sites, and an ensemble of Opus 4.7 judges compares desktop and mobile screenshots of each page.

  Over 100 trials, Claude Code with Opus 4.7 scored a mean reward of 0.80, against 0.98 for the oracle that feeds the ground truth back to itself. The task, not the attempt, decided most of the score: minimal layouts scored above 0.87, decorative ones below 0.75, and retro designs cost the agent about 0.18 against minimal ones.

• durable-execution

  2025 – Present

  durable-execution is a durable execution engine for running tasks and workflows resiliently, from a single function to a multi-step workflow. Tasks survive logic failures, process crashes and network errors, which makes it a good fit for AI workflows.

  It's published to npm as five packages, with storage for Convex and Drizzle and documentation.

• pdf-rag-embed-bench

  2025

  pdf-rag-embed-bench measures how well text and multimodal embedding models retrieve PDF pages: text embedders over parsed pages, multimodal embedders over page images, multi-vector embedders and combinations of them, with sparse retrieval and reranking.

  Multimodal embedders beat text ones, and combining a multimodal embedder with an image multi-vector one, sparse retrieval and a reranker reached 94.0% NDCG@5.

• microgpt

  2024

  microgpt is a small GPT implementation and training pipeline in PyTorch, which I built to understand attention, RoPE and multi-GPU training.

  It pretrains a GPT-2-class model in two stages: web data first, then three runs on smaller, high-quality data whose weights are averaged. The run took about four hours on 8 H200 GPUs, about $128 of GPU time.

  The 124M-parameter model peaked at 34.6% on HellaSwag, above OpenAI's GPT-2 (29.4%) and GPT-3 (33.7%) checkpoints of the same size.

• Programming tutoring system

  2014

  This was a research project under Dr Sumit Gulwani (Principal Researcher, Microsoft Research, Redmond) and Prof. Amey Karkare (Assistant Professor, IIT Kanpur).

  We built an application to assist in teaching introductory programming courses in the 1st year of college by automating assignment and feedback generation and solution-checking. The project was deployed in the 2014–15 academic year (with 400+ enrolled students) and was used by the students to practise programming assignments and get immediate feedback.

Education

Indian Institute of Technology Kanpur

B.Tech in Computer Science and Engineering, 2012 – 2016

CPI 9.6/10, Academic Excellence Awards 2013, 2014 and 2015

• All India Rank 196 in IIT-JEE 2012 among 500,000 candidates
• Indian National Mathematics Olympiad qualifier, 2010 and 2011
• KVPY fellowship, Government of India