Deedy Das

Deedy Das, also known as Debarghya Das, is a partner at Menlo Ventures focused on early-stage AI/ML, infrastructure, and enterprise software. He previously worked on Google Search and joined Glean’s founding engineering team, progressing into engineering management and creating and leading product work for Glean Assistant.[1][7]

Before Glean and Menlo, Das earned bachelor’s and master’s degrees in computer science at Cornell and worked at Facebook and Google. At Google, his work included query-understanding machine learning and cricket features across Search, Assistant, and Discover. That search experience supplied context for his later work on workplace retrieval.[1][8][5]

In his November 2019 essay, “Why I Left Google”, Das recalled preparing to join Waymo’s planning team in New York, then choosing a small, still-secret Palo Alto startup as a founding engineer. Later career records identify that startup as Glean. He wanted to learn to build systems beyond Google’s internal infrastructure: he had deployed services on Borg but was unsure whether he could scale a new product outside Google. His essay also discussed execution speed, professional growth, and uncertain compensation upside, while explicitly withholding another fundamental reason for leaving.[8][1]

At Glean, Das helped tackle information scattered across workplace tools. In his April 2023 Latent Space interview, he recalled Google’s internal Moma search as a reference point: employees could find a presentation even when they remembered its topic but not where it had been shared. He explained that robust SaaS APIs made documents and permissions accessible across services, enabling a comparable search experience outside Google.[5]

Das described Glean’s hybrid retrieval as combining traditional information-retrieval signals, synonyms, and acronym expansion with vector search. Ranking and personalization then helped determine which results mattered to the employee. His example was an engineer who was more likely to need documents authored by teammates than Salesforce documents. Workplace shorthand and team context were therefore part of relevance, alongside matching the query to document content.[5]ⁿ¹

He emphasized careful ranking work over technological novelty: “it's about the rigor and intellectual honesty that you put into tuning the ranking algorithm.” His account broke search into problems such as query understanding, synonyms, and document quality, each requiring sustained tuning.[5]

Das also argued that search alone gave employees too few reasons to return to a new tool. He described expanding Glean toward an employee portal, including Go Links—short, memorable addresses for workplace documents. This addressed a product problem beyond retrieval quality: employees might try a search tool once and forget it. His subsequent responsibilities included creating and leading product work for Glean Assistant.[5][1]

In the same 2023 interview, Das resisted treating ChatGPT as a universal replacement for search. He distinguished information needs such as fresh results and exploring multiple sources from questions suited to a generated answer. He explained retrieval-augmented generation as searching in the background and using the results to compose a response, and favored language models consulting tools such as search, calculators, and compilers for factual or deterministic work.[5]

Das joined Menlo as a principal in 2024 and became a partner in September 2025. In the firm’s welcome announcement, Tim Tully recalled a discussion about vector databases that continued for weeks through Twitter direct messages, with Das supplying papers and articles to support his arguments.[2][3][7]

Das helped launch the Anthology Fund, a $100 million Menlo–Anthropic initiative supporting early-stage AI companies, and participates in its investment team. Companies receiving investment are not required to build on Claude. Menlo’s September 2025 promotion announcement identified him as a Goodfire board member and an OpenRouter board observer.[12][3]

Das coauthored Menlo’s 2025 Mid-Year LLM Market Update with Tim Tully, Joff Redfern, and Derek Xiao. The report argued that builders prioritized model capability over price and usually upgraded within their existing provider. It reported that 66% of surveyed builders upgraded models within a provider, while 11% switched providers.[13]ⁿ²

The report estimated Anthropic’s share of production AI usage at 32%, compared with 25% for OpenAI and 20% for Google. These were dated, scale-weighted survey estimates of usage, not spending, based on 150 technical decision-makers at enterprises and startups. Menlo disclosed investments in companies discussed in the report.[13]ⁿ²

Alongside his company work, Das pursued programming experiments. In February 2024, he explored Neal Agarwal’s Infinite Craft, a generative game that combines pairs of elements. He began with exhaustive expansion, but reaching 1,020 elements left 519,690 possible pairs for the next expansion. He switched to sampled combinations, reused newly created elements, and sometimes injected target words—a strategy he called “beam search.” He reported synthesizing more than 10,000 elements and making 1,000 first discoveries, including Chana Masala.[9]

His mac_computer_use project adapts Anthropic’s computer-use demonstration to native macOS. Its documented design uses screen capture for observations and cliclick for keyboard and mouse actions, with a Streamlit interface and automatic screen-resolution scaling. The README warns that the agent has broad access to the Mac and that the demonstration depends on changing beta APIs.[10]

An earlier project, Deedy-Resume, carries a 2014 copyright and addresses fitting substantial information onto a single page. Das’s Apache 2.0-licensed XeTeX template uses an asymmetric two-column layout aimed particularly at undergraduate computer-science students. MacFonts and OpenFonts variants distinguish proprietary macOS fonts from redistributable alternatives; its documented limitations include overflowing onto a second page when a column becomes too long.[11]