sci2sci

Coding Agents need a proper workspace - so we made one! Announcing Xen - our agentic virtual monorepo orchestration. So what's the problem? Any more or less mature company has multiple projects - and every one of them knows that coordinating work across them is a time-burning mess. Big tech solves it with mono-repo structures - they allow you to work with all the company project like it's a single one. The problem? Well, your company is not Google (except the case you actually work in Google). We wanted to make it easier to complete the work across multiple software project as a single agentic coding session - all in all, with our dozen of projects, we face this issue every day. So our - quite brilliant - CTO, who worked first-hand with Google-scale infrastructure, made a scaled-down version of mono-repositories: keeping all their pros for agentic coding, and none of the headaches of migrating to them and supporting them after. If you're actively using tools like Claude code, or Codex, or our Parseltongue-CLI - and want to work across multiple software or research git repositories - this is an absolute must-have in your toolbox. I will share the link to our github in the comment, and if you want to check more technical explanation, dive into Valerii's announcement!

𝗪𝗲’𝘃𝗲 𝘀𝗼𝗹𝘃𝗲𝗱 𝗔𝗜 𝗺𝗲𝗺𝗼𝗿𝘆 𝗯𝘆 𝗿𝗲-𝘂𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝗵𝘂𝗺𝗮𝗻 𝗯𝗿𝗮𝗶𝗻 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲. 𝗟𝗲𝘁 𝗺𝗲 𝗲𝘅𝗽𝗹𝗮𝗶𝗻. The entire industry adds vector/graph databases to LLMs and calls it "memory." It's a category error. I spent years in neuroscience studying the molecular mechanisms of learning and memory. The biological brain has no store in a traditional sense. Not even stable weights like in artificial neural nets. Every stimulus triggers recalculation and reconfiguration across brain networks. Comparing real memory to RAG or a knowledge graph is like comparing a living expert to their notebook. Notebooks are useful, but they don't have intelligence or degrees of freedom. Memory is the behavior of continuous, global computation. And every time evolution has tried to preserve behavior across time, the medium has been the same: code. In genes, neurotransmitters, language, integrated circuits. 𝗖𝗼𝗱𝗲 𝗶𝘀 𝗱𝗮𝘁𝗮 𝘁𝗵𝗮𝘁 𝗰𝗼𝗻𝘁𝗮𝗶𝗻𝘀 𝗲𝗻𝗼𝘂𝗴𝗵 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝘁𝗼 𝗶𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁 𝗮𝗻𝗱 𝗲𝘅𝗲𝗰𝘂𝘁𝗲 𝗶𝘁𝘀𝗲𝗹𝗳. That's the principle behind Integrity Cortex. 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗮𝘀 𝗖𝗼𝗱𝗲. 𝗜𝗻𝘁𝗲𝗴𝗿𝗶𝘁𝘆 𝗖𝗼𝗿𝘁𝗲𝘅 combines stability of software and flexibility of a human nervous system. It transcribes every fact from your documents and data, every AI output, and connects atoms of knowledge into vast computational networks. Unlike RAG or knowledge graphs, it’s not passive - it actively processes information. While executing and updating the network, it validates that:  - all claims are grounded in evidence - derivations hold  - information across multiple sources remains consistent. Which means it catches LLM hallucinations. It also catches the inconsistencies that existed in your ground truth before AI ever touched it. Your enterprise data stops being a corpus and becomes a single running program whose shape changes every time new information arrives. Deterministic. Interpretable. Auditable by design. Example: when a primary endpoint is redefined, the cortex traces every dependency, every downstream report that now breaks coherence and needs an update - automatically. A week of document archaeology becomes minutes. __________________________________ We publicly announced Integrity Cortex at Genmab Data Day in Utrecht and Copenhagen last week - and couldn't ask for a better first audience, sharing the stage with Genmab’s internal teams and our colleagues from dbt Labs and Databricks. Big thanks to Kenneth Schwartz and the whole enterprise data team for the invitation and organization, and to all the participants for such engaged interest and great feedback! Ari K., Ashik Mirza, Mason Smith, Kevin Tsai, Brendan Kelly, Eko Baskoro Harimulyo, Elliott Bostelman, Frank Rebers, Kay Behnke MBA, Livia van der Pant, Lionel Blanchet, Maude Farrow, Miles Freeborn, Sanna Herrgård, Simon Tetens, Sumit Vishwambharan, Walter Muller

Data products aren't just a buzzword - they’re the engine driving the next generation of biopharma. 💡 It was a pleasure to join Genmab for their Data Days across both Utrecht and Copenhagen. Seeing every vertical of the organization - from R&D and clinical to manufacturing and operations - align on a unified data strategy was a masterclass in digital transformation. Shared the stage with the visionary teams at Genmab, Databricks, and dbt Labs to discuss how a robust data marketplace and specialized AI tooling are redefining what’s possible in our industry. It was also the perfect opportunity to present our latest product to this amazing group of peers. I’ll make a LinkedIn announcement soon - stay tuned!

𝗪𝗲 𝘀𝗼𝗹𝘃𝗲𝗱 𝗟𝗟𝗠 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀. Let me explain. Every data team in biopharma has had the same conversation: "We'd love to use LLMs on this, but we can't trust the output." We built the answer. Today we're open-sourcing #Parseltongue - a framework that doesn't ask LLMs to hallucinate less. It makes hallucination structurally impossible to hide. Everyone has been trying to make models smarter. Feed better context. But no amount of prompt engineering changes the underlying math - and math says that LLMs will lie whatever prompts you give. Instead of asking an LLM to summarize documents, Parseltongue asks it to encode them as a formal logic system. Every fact must cite a verbatim quote. Every conclusion must derive from stated premises. Every inconsistency in derivations is surfaced - the language checks itself. Two capabilities that matter if you run a data function in biopharma: → 𝗖𝗿𝗼𝘀𝘀-𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆 𝘃𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻: every LLM output is verified against the existing knowledge graph. If the model inverts a fact, quote verification fails - and the system red-flags every conclusion downstream that was built on top of it. In fact, it surfaces inconsistencies that were already present in your ground truth before AI ever touched it. → 𝗗𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝘆-𝗰𝗵𝗮𝗶𝗻 𝘁𝗿𝗮𝗰𝗸𝗶𝗻𝗴: full lineage between metadata records, down to the sentence. When source data changes, we detect what's now obsolete - with sentence-level precision. Static analysis, but for unstructured data. The LLM still does what it's good at - reading, extracting, understanding context. The formal engine does what LLMs fundamentally cannot - track provenance, enforce consistency and make information loss visible instead of silent. (If you haven't seen Eamonn M.'s recent post on compounding information loss across agent chains - there’s a reason it’s resonating. It’s really good at explaining why this problem is structural and not solvable by better prompts). Apache 2.0. Model-agnostic. 1500+ tests. The module for using in coding agents (like Claude Code) is coming with the next release. The link in comments. P.S. The most interesting comment thread challenging the approach: https://lnkd.in/dQBSbJJ9

𝗩𝗖𝘀 𝗰𝗮𝗹𝗹 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗴𝗿𝗮𝗽𝗵𝘀 𝗮 "𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻-𝗱𝗼𝗹𝗹𝗮𝗿 𝗼𝗽𝗽𝗼𝗿𝘁𝘂𝗻𝗶𝘁𝘆". We’ve been building them for the last two years - and the theory is missing a few critical pieces. Great piece by Jaya Gupta and Ashu Garg at Foundation Capital on the "context problem" in AI was published a month ago - link in comments. They argue that the next generation of software won't just capture what happened, but why (the "context graph"). At sci2sci, we couldn’t agree more. In biopharma, context is everything. Drug development cycles span 10-15 years; every decision made today - whether for a current pipeline or a new molecule - rests on a foundation of thousands of experiments conducted over a decade ago. Teams change, talent moves on; preserving institutional knowledge is daunting. But there’s a massive gap between VC theory and the reality of enterprise teams. The current thesis suggests we need "agents in the loop" recording every decision as it happens. Here is why it fails: 𝟭. 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗳𝗮𝘁𝗶𝗴𝘂𝗲: We’ve seen enough "dead" Confluence pages to know that asking users to document more is a losing battle. 𝟮. 𝗧𝗵𝗲 "𝗕𝗶𝗴 𝗕𝗿𝗼𝘁𝗵𝗲𝗿" 𝗲𝗳𝗳𝗲𝗰𝘁: Constant monitoring of every employee action creates more friction than it solves. 𝟯. 𝗧𝗵𝗲 𝗹𝗲𝗴𝗮𝗰𝘆 𝗴𝗮𝗽: Most importantly, the VC model is purely prospective. It ignores the legacy companies own. Our approach is different. Our customers have decades of data worth billions that they cannot afford to lose. We don’t wait for people to document things. We don’t "eavesdrop." Our engine acts as a digital archeologist. It scans siloed systems - cloud storages, network drives, lab software etc. - to automatically reconstruct the logical chain of a project - a context graph - even years after the fact. We recently helped a team return to a drug candidate that had been shelved for 3 years. Instead of months of "data forensics" to figure out why specific formulations were chosen/ discarded, our engine automatically connected the raw assay data to the original slide decks, lab notes etc. We restored the full decision-making context in minutes, not months. Two things the theory misses: The cross-platform reality: The article focuses on agents sitting mostly in one workflow (e.g. CRM). Real enterprise context is buried across silos. Ttrue context requires an engine that can bridge these platforms. Auditability as a requirement: In a regulated environment, context isn't "nice-to-have"; it’s a compliance necessity. We approach it by truth-grounding every AI decision with a full provenance chain, so that the reconstructed context holds up to the same scrutiny as the original study. Unlocking this "trillion-dollar opportunity" is not about recording the future - it’s about a way to make the context we already have usable, searchable and trustworthy. What do you think? #AI #Biopharma #DataGovernance #ContextGraphs #GenerativeAI #DigitalTransformation