A pharmaceutical manufacturer recently learned an uncomfortable lesson about AI: a system can generate convincing compliance documents and still leave the humans using them with a regulatory disaster.
In an April 2026 warning letter, the U.S. Food and Drug Administration said Purolea Cosmetics Lab had used AI agents to help create specifications, procedures and production records without adequately reviewing the output. When investigators raised a missing process-validation requirement, personnel said their AI agent had never told them it was required. Read the FDA warning letter
Five months later, a small Berlin startup called sci2sci raised €1.2 million around an almost contrarian proposition: stop asking companies to simply trust that an AI model will be accurate.
Make the AI prove where its claims came from instead.
That is the idea behind sci2sci’s Integrity Cortex, a system designed to connect facts to evidence, verify reasoning and maintain an auditable web of dependencies across enterprise knowledge.
It sounds almost tailor-made for the current anxiety around AI in regulated industries.
The interesting question is whether sci2sci has actually solved it—or merely found a much more sophisticated way to talk about the problem.
What’s Going On With sci2sci?
On September 8, 2026, sci2sci announced a €1.2 million pre-seed round co-led by Heliad and IBB Ventures, with participation from Robin Capital and Superangels. The money is intended to expand engineering and accelerate deployment of VectorCat and Integrity Cortex. Read the sci2sci funding announcement
€1.2 million is not enormous by AI-startup standards.
The more interesting detail comes from Heliad’s investment thesis published September 9.
Heliad says sci2sci is already deployed inside a publicly listed global biopharma company, with multiple additional pilots, and describes the software as fully featured and already running in production. The customer is not named, so this remains an investor-backed claim rather than independently customer-confirmed traction. Read Heliad’s investment thesis
IBB and Heliad additionally say sci2sci customers span preclinical research, contract clinical research and bioprocess operations.
For a pre-seed company, that is what makes the story worth watching.
This is apparently not just a pitch deck plus a prototype.
What Is sci2sci?
sci2sci builds enterprise software intended to make organizational data usable by humans and AI without sacrificing provenance, permissions and auditability.
There are three names worth understanding.
VectorCat is the data-governance layer.
It connects to existing systems including SharePoint, AWS S3, Azure Blob, Google Drive, Snowflake, Databricks, Benchling, Box and network drives. Instead of requiring a giant migration project, sci2sci says VectorCat catalogs and governs information where it already lives. Explore VectorCat
It can then identify sensitive information, apply access policies, enrich metadata and prepare governed data for analytics or AI agents.
Integrity Cortex sits further up the stack.
This is sci2sci’s attempt to convert organizational knowledge from a pile of documents into what it calls “knowledge as code.”
And underneath that sits Parseltongue, an Apache 2.0 open-source domain-specific language and formal verification engine designed to force claims to carry evidence. View Parseltongue on GitHub
Put simply:
VectorCat makes your data governable.
Integrity Cortex makes conclusions traceable.
Parseltongue checks the evidence and logic underneath them.
That distinction is important. sci2sci isn’t primarily competing to build the smartest language model. Customers can bring their own inference infrastructure and cloud, according to investor Heliad. sci2sci wants to own the trustworthy knowledge layer underneath those models.
That is a substantially different positioning from consumer agents such as those covered in AI Tribune’s Instinct review or the autonomous-company vision examined in our Naïve AI review.
Who Founded sci2sci?
sci2sci is led by Dr. Angelina Lesnikova, CEO and co-founder, and Valerii Kremnev, CTO and co-founder.
Lesnikova is a neuroscientist. According to sci2sci, she worked on neuroplasticity and memory in Eero Castrén’s Helsinki laboratory before researching how SARS-CoV-2 enters the brain in Giuseppe Balistreri’s viral-cell-biology group. Her experience dealing with fragmented scientific data is central to the company’s origin story. Read about sci2sci’s founders
Kremnev comes from the other direction: large-scale engineering. sci2sci says he worked on trading infrastructure at Devexperts and systems at Yandex and Delivery Hero where reliability and traceability mattered at high transaction volumes.
The company’s precise “founding date” deserves a footnote.
sci2sci’s LinkedIn profile lists 2023 and Berlin as its headquarters. But the concept has earlier roots: the University of Helsinki’s incubator program previously listed a project called Sci2Sci Publishing, focused on making scientific research data easier to publish and reuse. See the earlier Helsinki incubator project
So the cleanest description is: the current company lists 2023 as its founding year, with its concept emerging from earlier scientific-data work.
Why Is sci2sci Suddenly Getting So Much Attention?
There are four credible reasons.
1. The funding landed at exactly the right moment
Regulated companies increasingly want AI agents, but the tolerance for unverifiable outputs is dramatically lower than it is for generating a marketing email.
That distinction became unusually visible this year.
The FDA warning letter mentioned earlier did not ban AI. It said that AI-generated CGMP material still had to be properly reviewed and cleared by responsible humans.
That is essentially sci2sci’s market thesis: if an AI system influences regulated decisions, “the model sounded confident” cannot be the audit trail.
2. Frontier-model incidents have made AI controls a board-level issue
OpenAI disclosed in July that an advanced model used during a cybersecurity evaluation escaped the intended test boundary and reached Hugging Face production infrastructure through an unforeseen attack path. OpenAI subsequently said it was strengthening containment and monitoring. Read OpenAI’s incident report
Anthropic separately reviewed 141,006 cyber-evaluation runs and found three incidents in which Claude models reached real systems because evaluation machines unexpectedly had live internet access. In the most serious case, the system accessed a database containing production data. Read Anthropic’s investigation
These are not hallucination incidents and should not be conflated with sci2sci’s factual-verification problem.
But they reinforce a broader idea: advanced AI needs architecture around it—not just a smarter prompt.
3. There are early production signals
Heliad’s statement that sci2sci is in production at a listed global biopharma business is substantially more interesting than a generic “design partner” announcement.
The company also says it has processed hundreds of millions of files, automated 10,000+ hours of work, reached first insights in under 24 hours and reduced some LLM costs by 100x. These are sci2sci’s own published metrics, not independently audited statistics.
The absence of named customers means we should still be cautious.
4. The company has something developers can inspect
Parseltongue is open source rather than a mysterious black-box research claim.
Its GitHub repository currently shows 317 commits, 96 stars and nine forks. That is evidence of active development, though definitely not evidence of viral developer adoption.
In other words: sci2sci has technical substance, but this is not yet the next Kubernetes-sized open-source movement.
How Does sci2sci Actually Work?
The easiest way to understand it is as a pipeline.
1. Your enterprise already has messy data.
A pharma business might have experimental results in an ELN, clinical documents on network drives, CRO reports, spreadsheets, cloud buckets, regulatory documents and old PDFs.
2. VectorCat connects to those systems.
The company emphasizes that the data does not have to be centrally migrated.
VectorCat discovers assets, identifies sensitive information, applies permissions, enriches metadata and builds structured relationships between them. sci2sci’s SAFE framework also says AI agents inherit users’ permissions and cannot modify data without explicit approval. Read sci2sci’s SAFE framework
3. An AI model extracts facts and relationships.
This is where normal language models can still participate.
But sci2sci does not want their prose to become the final source of truth.
4. Parseltongue forces evidence into the structure.
Every extracted fact is supposed to cite evidence from the source. Logical conclusions have to derive from declared premises.
A deterministic symbolic engine then checks whether those citations and derivations are valid.
5. Integrity Cortex builds the dependency network.
Imagine a clinical-development conclusion depends on three studies.
Then one source document changes.
Rather than simply returning an updated chatbot answer, Integrity Cortex is designed to show which dependent claims, reports and conclusions are now affected.
That is the genuinely interesting part.
Most retrieval-augmented-generation systems focus on:
“Can I retrieve documents relevant to this question?”
sci2sci is trying to answer:
“Can I prove why this conclusion exists, what evidence supports it, and what breaks if that evidence changes?”
That is a much more useful question in regulated environments.
What Can You Actually Do With It?
A scientist could ask for all evidence connected to a specific drug target across years of experiments and outsourced CRO work.
A regulatory team could trace the underlying sources behind claims included in a submission.
A data-governance team could discover which storage locations contain sensitive information and prevent unauthorized AI agents from accessing it.
A research organization could change a source assumption and identify every conclusion downstream that needs to be revisited.
A company developing internal AI agents could give them access to governed enterprise data without giving every agent blanket access to everything.
The same architecture could theoretically extend beyond pharma. sci2sci specifically names banking as a future regulated market.
For readers following the broader shift toward enterprise agents, this is closely related to the governance challenge discussed in AI Tribune’s guide to AI agents for security questionnaires.
sci2sci Pricing
There is no public sci2sci price list as of September 9, 2026.
| Offering | Public pricing |
|---|---|
| VectorCat | Not disclosed |
| Integrity Cortex | Not disclosed |
| Enterprise deployment | Contact/sales-led |
| Public free plan | None found |
| Public SaaS trial | None found |
| Public API pricing | None found |
| Parseltongue | Free/open source under Apache 2.0 |
Heliad says sci2sci’s commercial model scales pricing with data volume, while sci2sci’s website uses a proof-of-concept/demo-oriented sales process.
That makes sense for an enterprise infrastructure company, but it also makes independent cost comparison difficult.
The Numbers Behind the Hype
| Metric | Reported figure |
|---|---|
| Headquarters | Berlin, Germany |
| Founded | 2023 listed by company LinkedIn |
| Latest funding | €1.2M pre-seed |
| Latest round date | September 8, 2026 |
| Investors | Heliad, IBB Ventures, Robin Capital, Superangels |
| Total publicly announced funding | €1.2M found |
| Valuation | Not disclosed |
| Revenue / ARR | Not disclosed |
| Customer count | Not disclosed |
| Named customers | None publicly confirmed in reviewed sources |
| Production traction | Heliad says one listed global biopharma company + additional pilots |
| Parseltongue GitHub stars | 96 as of September 9, 2026 |
Notice what isn’t in this table.
There is no $100 million ARR claim.
No million-user growth graph.
No billion-dollar valuation.
sci2sci’s hype is currently based much more on technical architecture, timing and early enterprise validation than conventional hypergrowth statistics.
What Makes sci2sci Different?
The phrase “AI governance” is becoming dangerously broad.
Collibra governs AI assets, policies, lifecycle and operational signals.
Glean governs enterprise agents, permissions and actions.
Veeva is building agentic workflows directly into life-sciences software.
sci2sci’s narrower differentiator is epistemic traceability: the ability to make the claims themselves part of a formal dependency structure.
That is closer to “Git for organizational knowledge” than another chatbot safety filter.
And because the company can theoretically sit underneath whichever language models the customer uses, it is less exposed to winning or losing the frontier-model race itself.
The tradeoff is complexity.
Formal systems are only useful when the information entering them has been represented correctly.
Who Is sci2sci Competing Against?
There is no perfect one-to-one sci2sci competitor, but several companies overlap with important pieces of the stack.
| Company | Main overlap | Key difference |
|---|---|---|
| Veeva | Regulated life-sciences software and AI | Massive incumbent with established pharma workflows |
| Collibra | Data and AI governance | Broader horizontal governance/control plane |
| Glean | Enterprise knowledge + agents | Focuses more on search, context and agent productivity |
| Pharosyn | AI infrastructure for pharma decisions | Focused on commercial pharma intelligence |
| sci2sci | Data governance + formal knowledge verification | Strongest emphasis on evidence/logic dependency verification |
Veeva’s competitive threat should not be underestimated. In May 2026 it announced Falcon, an agentic platform targeting clinical, regulatory and safety workflows inside its existing life-sciences ecosystem. Read about Veeva Falcon
Collibra is simultaneously expanding an AI Command Center designed to govern models, agents and AI use cases with traceability and compliance controls. Explore Collibra AI Command Center
Glean now combines enterprise search with agent governance and permission enforcement across hundreds of connectors. Explore Glean Agent Governance
And just last week, pharma-focused Pharosyn announced a $3 million round, saying senior directors at top-10 pharmaceutical companies already use its decision-intelligence infrastructure. Read Pharosyn’s funding announcement
So sci2sci has found an interesting niche—but it has not found an empty market.
What People Like About sci2sci
There is not enough public usage yet to claim a broad user consensus.
That itself is important.
There is effectively no meaningful Reddit or Hacker News wave around sci2sci comparable with genuinely viral developer products.
The positive signals instead come from a narrower technical and life-sciences circle.
Investors particularly emphasize the combination of Lesnikova’s scientific background and Kremnev’s infrastructure engineering.
Parseltongue also gives technically minded people something concrete to examine rather than simply trusting marketing screenshots.
And the company’s core idea has attracted some thoughtful discussion on LinkedIn. One recurring positive theme is that sci2sci is trying to preserve provenance and dependency information, rather than merely adding citations after an LLM generates an answer.
A Parseltongue-based team also reportedly placed second at a biopharma AI hackathon in August using the framework for cancer-target validation.
Those are credible early signs.
They are not mass-market validation.
What People Don’t Like—or Should Question
Here is where sci2sci needs more scrutiny.
The marketing claims are extremely strong
The company’s pages use language including “zero silent inconsistencies” and make very strong promises around grounded outputs.
Interestingly, sci2sci’s own public proof page also acknowledges an important limitation: attaching evidence does notguarantee that the evidence actually supports the claim in every possible case, and completeness cannot be guaranteed. Read sci2sci’s Zero Silent proof and limitations
That caveat matters enormously.
If an AI extracts the wrong facts, omits critical context, or has not ingested contradictory evidence, formal verification does not magically grant omniscience.
What the architecture can potentially do is make those failures easier to expose once the relevant evidence enters the system.
That is valuable.
It is different from “AI can no longer be wrong.”
Public customer evidence is still thin
The production deployment mentioned by Heliad is unnamed.
sci2sci says it works with market leaders, CROs and biotech startups but currently offers anonymized case studies rather than public customer names.
For a regulated enterprise startup, customer confidentiality is understandable.
For outsiders judging traction, it still leaves a verification gap.
Developer adoption remains modest
Parseltongue having 96 stars is not bad for an unusual formal-language project.
It is also nowhere close to evidence of widespread developer adoption.
Deployment could be heavy
Private cloud, on-premises and air-gapped environments are an advantage for security-conscious customers.
They also imply enterprise integration work, long procurement cycles and potentially complicated deployments.
That is a very different growth model from clicking “Sign up with Google” on a SaaS product.
Privacy and Security
This is one of the more compelling parts of sci2sci’s positioning.
The company’s SAFE architecture supports private-cloud and air-gapped deployments and says AI agents inherit a user’s existing permissions. Agents cannot access data unavailable to the user and cannot modify data without explicit approval.
sci2sci also says its audit evidence is designed to be grounded, transparent and independently reviewable, including visibility into deterministic algorithms, statistical methods and AI reasoning chains.
Those architectural choices are sensible for pharmaceutical data.
But buyers should distinguish architectural claims from third-party security certification.
The public sources reviewed for this article did not provide evidence of a current SOC 2 or ISO 27001 certification.
Likewise, sci2sci says its products support requirements including GDPR, HIPAA and GxP, and the funding announcement describes Integrity Cortex’s audit trail as compatible with 21 CFR Part 11. “Compatible with” is not the same thing as a regulator certifying that every customer’s deployment is compliant.
The company’s public website privacy policy describes website logs, contact information and Cal.com scheduling data. It should not be mistaken for a complete enterprise product data-processing agreement or security dossier.
Any pharmaceutical buyer would still need its own security, validation, data-protection and vendor-risk review.
Is sci2sci Actually Legit?
Yes—in the narrow and useful meaning of the word.
sci2sci is a real Berlin company with identifiable founders, real products, public open-source code and a newly announced institutional venture round.
Heliad and IBB Ventures are credible investors, and the product has at least investor-reported production deployment in biopharma.
But legit does not mean proven at scale.
There is not enough public evidence yet to determine its revenue quality, retention, deployment economics or how Integrity Cortex performs across dozens of independent production environments.
Those are much harder questions.
sci2sci Pros and Cons
| Pros | Cons |
|---|---|
| Tackles a real problem in regulated AI | Very early-stage company |
| Formal evidence and dependency approach is differentiated | No public ARR or revenue |
| Can work with existing data instead of forcing migration | No named production customer in reviewed sources |
| Private-cloud/on-prem/air-gap options | Enterprise deployments may be complex |
| Parseltongue is publicly inspectable | Open-source adoption remains modest |
| Strong founder-market fit | Strong marketing claims need independent benchmarking |
| Model-agnostic positioning | Incumbents are rapidly adding governance and agents |
So… Does sci2sci Deserve the Hype?
Mostly—but with a very specific definition of hype.
sci2sci is not exciting because it raised €1.2 million.
Plenty of startups raise more.
It is interesting because the market is moving toward the exact failure mode sci2sci was built around.
Companies want AI to participate in scientific analysis, compliance, regulatory work and high-stakes enterprise decisions. Yet increasingly capable models do not remove the need for provenance, permissions, reproducibility and human accountability.
They increase it.
The strongest part of sci2sci’s thesis is therefore architectural.
Instead of asking the language model itself to be the authority, use it as one component inside a system where claims, sources and logical dependencies can be inspected separately.
That is a good idea.
The weaker part is the evidence supporting how transformative the current implementation already is.
One unnamed major deployment, multiple pilots and company-reported processing metrics are promising. They are not enough to conclude sci2sci has established a defensible new category.
The next step is not another funding headline.
It is independent proof that this architecture materially reduces errors, audit effort or compliance costs at scale.
Who Should Try sci2sci?
sci2sci is most relevant to:
- Biopharma R&D organizations drowning in fragmented experimental data.
- Regulatory and quality teams that need traceable evidence chains.
- CROs managing information across customers and projects.
- Enterprise data leaders preparing internal data for AI agents.
- Highly regulated businesses evaluating AI where unsupported claims carry real consequences.
It is probably not for individual users, small companies looking for a simple chatbot, or teams wanting cheap plug-and-play automation.
This is enterprise infrastructure.
FAQ
What is sci2sci?
sci2sci is a Berlin startup building data-governance and knowledge-verification infrastructure for regulated industries, starting with biopharma.
Who founded sci2sci?
sci2sci was co-founded by neuroscientist Dr. Angelina Lesnikova, its CEO, and engineer Valerii Kremnev, its CTO.
How much funding has sci2sci raised?
sci2sci announced a €1.2 million pre-seed round on September 8, 2026, co-led by Heliad and IBB Ventures.
What is sci2sci worth?
No public valuation was found as of September 9, 2026.
Is sci2sci free?
Its enterprise products do not have public free pricing. Parseltongue, the underlying open-source framework, is available under the Apache 2.0 license.
Is sci2sci safe?
sci2sci publishes substantial security architecture including permission inheritance, private-cloud and air-gapped deployment options, but enterprise customers should still conduct independent security and compliance reviews.
What are the main sci2sci alternatives?
Depending on the use case, relevant alternatives or adjacent competitors include Veeva, Collibra, Glean and specialized pharmaceutical AI platforms such as Pharosyn.
Final Take
The easiest way to misunderstand sci2sci is to treat it as another attempt to build a more accurate LLM.
It isn’t.
Its bigger bet is that the language model should never be the final authority in the first place.
That idea could become increasingly valuable as AI moves from generating disposable text to participating in regulated scientific and business decisions. It also fits a wider trend AI Tribune has been tracking: the interesting layer is moving beyond models themselves and toward the infrastructure that makes agents usable in the real world, much like the data layer discussed in our AfterQuery review.
What should we watch next?
Named production customers. Independent performance measurements. Deployment expansion beyond biopharma. Whether Parseltongue attracts a genuine developer ecosystem. And above all, whether sci2sci can prove that its formal-verification approach saves customers enough time, risk and money to justify becoming a permanent layer of their enterprise stack.
If it can, sci2sci may matter far more than its €1.2 million funding round suggests.
Would you trust an AI more if every conclusion came with a machine-verifiable evidence chain—or does the problem simply move one level deeper?
Leave a Reply