Nettle is an autonomous data platform that takes away grunt work from your data teams. We ingest your enterprise data (legacy or otherwise), structure it, classify it, join on it, break it down and build models over it. We are optimized for rapid transformation of data to insights and business value, often producing business impact in less than 30 minutes. A typical user for Nettle saves hours of time per usecase and is working with a stable, governed platform rather than an ad-hoc set of LLM prompts and scripts. Agents query prepared data instead of reading raw files, so they spend fewer tokens. Nettle offers more than 70 MCP tools and can be controlled with prompts from your favorite model provider and harness. Our GRC infrastructure provides visibility over AI costs, token use and compute. The system has no external dependencies other than calls to LLM models and can be run airgapped with local models if needed. The system only reports telemetry for monitoring and can be run without it if needed for high security applications.

Divye left Pinterest on April 19, 2024, a Friday and started full time on Nettle on April 21, 2024 - the next Monday. He spent the remaining half of 2024 and early 2025 on internal R&D for Nettle with part-time ops support. In February 2025, Divye funded an R&D group with high-risk capital from his savings through a family trust. The R&D group was active from February to November 2025 and all early activities were executed under Flinking LLC, a California Limited Liability Company. Flinking LLC was reorganized as a subsidiary of Nettle Holdings Limited on July 29, 2025 and the first capital injection of $800K was provided by friends and family in September-November 2025. Divye considers November 2025 as the founding month of the company as the company transitioned from whitepapers to a commitment to build the underlying system. Today, Nettle is the product and brand of Nettle Holdings Limited, the company's legal entity, and Flinking LLC (California) is its wholly owned operating subsidiary.

Between Feb 2025 - May 2025, an initial 10 person R&D group ran a 10 week research sprint with daily and then alternate day meetings to clarify and identify the key structural gaps that were causing problems in real-world data systems, including data representation, data dynamics and control and governance. A terminology bridge was built between Computer Science, Physics and Math so that ideas could be lifted between these domains. Between May 2025 - November 2025, the R&D team consolidated the key ideas and 5 core contributors located key areas of the literature relevant to Nettle’s problem domain in both physics and math. This allowed the R&D team to roll off into an advisory role. Between November 2025 - February 2026, the core R&D ideas from physics and math were converted to computer science formulations and the key ideas of Nettle were crystallized in 30 hours of video over the course of 9 days in a “Crunch Week” in February 2026. The crystallization in the Feb 2026 Crunch Week enabled the company to start converting the key ideas to Computer Science and code starting March 2026, when the initial engineering team starting moving to a full time role on the project. An initial version of the product went live around July/August 2026 and is being actively tested by potential customers. The R&D phase is now considered largely complete, though smaller problems continue to be resolved as implementation proceeds.

Resolving the R&D issues, transitioning from math and physics conceptual structures to implementable computer science and building the system and testing it with design partners were the promised milestones of the Nov 2025 friends and family capital raise. These have now been largely achieved. We have an active product live and accessible from our website, an early GTM motion and multiple pilot evaluations in flight. The product is in an early stage of commercialization, as is expected for the company stage and the company is raising capital to increase team bandwidth, speed up GTM and continue product development and commercialization.

Companies at the Series B+ stage and beyond are the ones that typically need data platforms. They have the large teams of 10+ data engineers, data scientists and analysts. When building a solution in the domain, Nettle has to solve exactly the same problems of a much larger company at the pre-seed stage. Our team size is appropriate purely because a) the problem space demands it, b) the implementation can be parallelized and hiring appropriately actually increases ROI and c) the team is able to absorb larger projects and increase revenue without impacting the core product build during the execution period. The general market in this space operates with similarly sized teams and a smaller team impacts our ability to execute at pace and capture the market.

The raised capital is expected to create a 2x 8-person team structure (7 team members + 1 manager) on both engineering and GTM and fractional support on G&A functions and sufficient headroom for cloud compute and AI costs. The structure efficiently uses management bandwidth, provides enough engineering heft for the team to absorb larger clients, larger per-client variances, product internationalization, data migration and ML model customization use cases. The GTM team, in turn, also can run a global GTM strategy from day 1 and target higher ACV clients. The team is expected to scale on a ramped schedule as the client base builds up and not aggressively ahead of actual closed deals. Since services costs per deal are likely to be the same whether it’s a larger client or a smaller one, a better staffed GTM & Eng team with the ability to handle larger enterprise deals is likely to yield a higher revenue outcome at the next fundraising milestone rather than a thinly staffed one. The larger teams, name brand VCs and a stable team structure also provides the necessary execution trust as buyers in the data space buy the perception of trust, safety and quality more than they buy the hard metrics of the product itself. By the end of 24 months, Nettle should demonstrate a commercial solution generating revenue of up to $1M with a healthy pipeline of clients and a clear understanding of why people are buying the product and the value they’re deriving from it. Nettle is currently at a “pre-Seed” post-product and pre-revenue stage and these are healthy milestones for the company to target in time for its “Seed” fundraise.

As of Sept, the product is in public beta. Customers can try the product online with publicly accessible data at no cost. After they gain confidence, we sign contracts and create a private workspace for them.

The design partner program is a 1-2 quarter engagement with a prospective client where we address a business problem for them using Nettle at a heavily subsidized cost to customer (in some cases $0). During this period the partner agrees to help improve the product by sharing data and feedback for the duration of the engagement. Since beginning the program in last August 2026, we are in the contract execution stage with 3 prospective partners. Our current partner pipeline is about 10 more such partners. The design partnerships program provisions 10 such partners to help validate the product and generate traction.

Our outbound campaign (cold email and phone) has reached more than 2,000 leads (as of October 2026). That has produced 25 interested prospects and 7 completed or scheduled demos. Three of those prospects are in active discussions to join our design partner program, and one has already signalled willingness to pay for a monthly subscription. Conversion through the funnel is up roughly 30% versus our first month, driven by ongoing refinement of our ICP and messaging as we learn from each prospect conversation.

High confidence ingest to insights in 10 minutes.

Nettle can run alongside or on top of existing data stack.

Based on our design partnership engagements, setting up Nettle for a new customer does not take more than a day. What is time consuming is the customer due diligence and contractual steps that take longer to the tune of 4-6 weeks. We can considerably shorten this with all out security and compliance certifications and postures in place.

Nettle is autonomous and data agnostic by design and based on our generalised data model that allows for Nettle to work just as well on previously unseen data. There maybe minor customisations to suit a particular use case, but our core product does not change.

Nettle is an autonomous data processing platform. It's allocation comes out of the SWE savings it provides.

The current round is pre-seed at $3M. It pays for the next 2 years of expenses most of which are on team expansion, product maturity and early sales traction.

As of October 2026, Nettle has raised an $800k F&F round that will last us till December 2026 (inclusive).

The current monthly burn is about 80k (as of October 2026). We hired GTM in order to secure design partners.

Yes this is still our view for a $1.3B exit.

Across 125 completed datasets on the live platform (as of October 2026), active processing time is: median 11m 10s, p75 15m 23s, p90 20m 45s, fastest 22s, slowest 70m 37s. Wall clock including queue time is median 14m 04s and p90 31m 32s once you exclude nine stalled runs. By source, Kaggle and Hugging Face land around 10 minutes, direct URL and local uploads closer to 20. Size matters less than expected: under 10k rows is 11m 16s median, over 10M rows is 20m 43s.

As of October 2026, the largest processed is 37,290,032 rows across 6 tables, finishing in 18m 42s. The widest is 103 tables at 19.4M rows in 20m 43s. Everything runs single-node today; there is no distributed mode, and the upload ceiling is 20 GB per submission. Those are tested limits rather than design limits.

95% accuracy (as of October 2026). When a model is wrong, the use can find out on the dataset metastore which allows classification updates by hand as well.

Column classifications can be reviewed and overridden inline, with filters for low-confidence and personal-data columns so you can go straight to what the model was unsure about. Datasets are soft-deleted with a restore path rather than destroyed. Every ingest records who triggered it, whether it came from the UI, the API or an AI agent, plus IP, location and timestamp, and every job keeps a full log stream with per-call LLM cost.

Full SQL against the complete dataset runs in the browser, and everything exports in open formats: DuckDB, Parquet, Iceberg, SQLite, plus the report as a document. The whole platform is also exposed over MCP, so an engineer can drive it from their own tooling. There is no Python execution, and no git or dbt integration today. Generated transformations cannot be hand-edited; you override the classifications that feed them, or take the export and work downstream.

At this point the formal model changes the way we implement classifications, making it dataset agnostic and more durable for unseen datasets.

The table and column classification modules implement the model. Implementation of the Silver layer as WIP also use the model for 3rd Normal Form representations. Addition to this, all downstream modules also read from the upstream formal models.

There are technical notes for our internal implementations of the formal model. However this will remain internal as trade secret. There are opporutnities where we will publish parts of the our implementation as marketing material or technical expositions.

Kashyap owns the core engine. Math advisers come in for alignment and validation of model.

Trade secrets that are not yet registered since work on implementation is ongoing.

Shipped today: ingest from eight sources, schema and statistical profiling, classification and tagging, the quality report, Nia, SQL, search, drift comparison, the storage depot and the MCP server. In beta and labelled as such in the product: Neal, Ont and the EDA canvas. Planned and shown as "Soon": Joinery for cross-table joins, Refinery for cleaning, Interpolate for time-series gap-fill, Predict for model opportunities, and the Silver layer. Beyond that the deck's own roadmap has prescriptive insights, coordination, governance, and access control.

We consume 5 models (as of October 2026) and that number is liable to increase. We consume from OpenAI, Google and Anthropic Models with some open weight models that come from fireworks. The data that is sent to them is short samples of schema and small sets of data rows that help accelerate initial data ingest phases. Providers are interchangeable with minor variations in overall accuracy. We maintain all frontier models and open weight models to ensure we can work with those models based on customer security posture and provider agreements.

The compliance road map is to maintain all the following compliances by Q2 2027. SOC2 Type 2, HIPPA, PCI DSS, ISO 27001, ISO 42001, FedRAMP, StateRAMP, TXRAMP and others as needed.

Customer data is deleted by user action by demoting to a quarantine for 72 hours for recovery if needed. Following this data is purged and leaves only a shallow tombstone to indicate data was ingested in the past. System also supports hard delete actions available on MCP and API routes (not available on UI today).

Nettle deploys a community version demonstrating the full capabilities of the platform on public datasets. When deployed for customers, Nettle is deployed on VPC. That workspace lives behind and OIDC specific auth and external connections are no available unless requested.

VPC and On prem as desired by customer. Air gapped is possible with appropriate LLM provider choices. Air gapped is not offered today. VPC and On prem are the options available to customers today.

k8s availability targets are 99.9%. Current status is maintained on status.nettle.llc and cluster uptime as of October 2026 is 99.092%

PII fields are detected and allow functionality to treat them. Right now these are not implemented since most of the data ingested is public data on community edition of Nettle.

Penetration testing is scheduled post fund raise in Q1, 2027.

1. Sequence Genie builds context from pipelines, queries and dashboards you already have. On a source system nobody has modelled, none of that exists yet. We work the week before the lakehouse, and export to Parquet and Iceberg so the output lands in Unity Catalog. 2. A different kind of check Lakeflow expectations test rules you wrote. We test relationships between columns nobody thought to check. On a recent customer file the churn flag disagreed with actual purchase behaviour for 20% of the base, 2,153 marked churned still buying. That column is fully populated, so nothing in your stack flags it. 3. Confidence is the risk Land a quietly broken column in Databricks and Genie will answer questions from it fluently and wrongly. The better your AI layer, the more convincing the wrong answer. We tell you which columns can carry the weight before you build on them.

1. Sequence Semantic View Autopilot works on Snowflake tables. It reads your schema, query history and BI assets, which is the right approach once data has landed and someone has decided what it means. We work on the estate before any of that exists, and emit Snowflake semantic view DDL so the output lands in your account rather than beside it. 2. A different kind of check Autopilot infers structure. It does not test whether your values agree with each other. On a recent customer file, 25% of records had a last purchase dated before the signup date, and the churn flag was wrong for 20% of the base. Every one of those columns is fully populated, so completeness checks and schema inference both pass them clean. 3. Snowflake's own advice makes the point Their field guidance for Autopilot is to validate the output with subject matter experts before deployment. That validation step is the work. We do it automatically and ship the evidence query with every finding, so your SME is checking an answer rather than building one.

Fivetran: complement, different layer. Fivetran moves tables and assumes you already know where they are going and what they mean. Nettle is what tells you which of the four CUSTOMER tables is authoritative. No overlap at all. Airbyte: complement, already in our stack. We use Airbyte for keep-fresh syncs on recurring sources. It is a component we run, not a competitor. dbt Copilot: complement, adjacent. dbt Copilot helps you write and document transformations inside a project that already exists. Nettle produces the understanding that tells you which transformations are worth writing. The obvious join is emitting dbt models as an output format, which is the right roadmap item and is not built today. Fabric Copilot: complement by sequence. Same shape as Snowflake and Databricks. It works on what is already in OneLake and modelled; we work on the estate before that decision has been made. Atlan: Atlan harvests metadata from systems that are already instrumented. We generate it for systems that are not.

Porter assumed settled boundaries; this category is about three months old, created when OpenAI, Anthropic and IBM committed over 5.5 billion dollars to forward-deployed engineering. The useful frame is a stack, and where each competitor's automation starts. Fivetran moves tables but will not tell you which of four CUSTOMER tables is authoritative. Snowflake and Databricks automate semantic modelling for data already inside their platform and will never optimise for data that has not landed, because their revenue depends on it landing. Catalogs need instrumented systems; a legacy estate has none. That leaves the seam we work in, day one on an unmodelled estate, structurally unattractive to all of them. Where we lose: anywhere the data is already landed and modelled, against Palantir at the top end, and on connectors, where we support files and URLs while competitors offer databases, CDC and on-prem.

S3 yes, as a read source for public objects. Iceberg yes, but as an output: we write the processed data out as Iceberg, Parquet and DuckDB, we don't read an existing Iceberg table. Snowflake, Databricks and BigQuery have no connector today, though you can point Nettle at a file exported from any of them and it ingests like any other URL. Nettle doesn't write back into any platform; everything it produces lands in its own storage.

Nothing is held in a proprietary format. The processed dataset downloads as a single DuckDB file, and the depot also writes Parquet/Iceberg, SQLite and a search index, all standard. The report exports as a document. If you stopped using Nettle tomorrow you'd keep the data and the inferred model in formats any tool can open, so leaving is a download rather than a migration.

Engineering is four people: Nitin Shukla, Mayank Chandravanshi, Alexey Sanko and Kashyap Kale. Arnab Ganguly, Stephen Stroer and Mohit Kumar sit on product and also commit, and Divye writes code alongside them, so the number of people shipping is larger than the engineering headcount.

All of it checks out and is easy to link. Flink Forward Seattle 2023 speaker page, two Pinterest Engineering posts on Medium ("Tuning Flink Clusters for Stability and Efficiency" and "Unified Flink Source at Pinterest"), a Google Scholar profile, a Sessionize speaker profile and the divyekapoor GitHub account. The public record describes Divye as Senior Staff Software Engineer at Pinterest leading their Flink-based stream processing platform, with Google and LinkedIn before that, and the Flink Forward talk claims 50 to 90 percent per-job cost reduction.

Dataset bhaskerpaul/sales-and-marketing-dataset - public on Kaggle. 15,000 customer records, 30 columns, one table. Nothing pre-arranged: no schema supplied, no data dictionary, no configuration. Steps to replay Connect Nettle as an MCP server to any agent: claude mcp add --transport http nettle https://app.nettle.llc/mcp Ask it to import bhaskerpaul/sales-and-marketing-dataset Wait for ready, then ask questions in plain English Or do the same in the browser: app.nettle.llc → Ingest Data → From URL → paste the Kaggle URL. Same pipeline, same artefacts. Timings (measured, 25 Aug 2026) Stage Duration Import → raw tables loaded, profiled, classified 2m 08s Gold layer built (162 pre-aggregated tables) 11.0s Executive report generated 2m 00s Total, API call to finished report 4m 20s The report records its own compute: 49.2s wall-clock, 18 SQL queries, 9 LLM calls on claude-sonnet-4-5, 9,657 tokens, $0.0669 measured API spend. Outputs DuckDB database, 5.52 MB - downloadable at /api/v1/datasets/d_02d949ce/exports/duckdb 162 Gold pre-aggregate tables, plus SQLite and a Whoosh full-text index An 8-section executive report: topology, join strategy, quality analysis, ML/dimensional modelling, query optimisation, cost metrics, appendix, and a fit-for-purpose verdict — with ER diagram and charts, downloadable Engine versions stamped on the batch for reproducibility: DuckDB 1.5.5, PyArrow 25.0.1, PyIceberg 0.11.0, Python 3.13.15, Whoosh 2.7.4 Validation results Two layers, and they deliberately disagree - which is the point. Structural: 99% overall (grade A), 99.3% completeness. Referential integrity returns N/A - no FK relationships to validate, correctly, because it's a single table. Semantic, from agent queries against the same data: 3,762 records (25.1%) have a last purchase date preceding the signup date - $1,835,729 of recorded revenue, 25.1% of the $7.3M total The churn flag disagrees with purchase behaviour on 3,026 customers (20%): 2,153 flagged churned are still buying ($2.67M lifetime value), 873 flagged active have stopped ($1.07M). Only 145 genuinely churned - 1%, against a reported 15.3% Each finding returns with the SQL that produced it. Individual checks run in 14–55 seconds at $0.015–$0.064 each. Link https://app.nettle.llc/session/d_02d949ce/catalog On replayability, precisely: the measured facts reproduce exactly — same row counts, same 3,762, same dollar figures, because they come from SQL. The report's prose and the agent's phrasing are model-generated and will vary between runs. If you want byte-identical output, the DuckDB export and the report are downloadable artefacts from this run.

Everything, and they don't need an account. app.nettle.llc is open, so in half an hour someone can paste a Kaggle URL or upload a file, get a finished report back in a few minutes with the schema, quality scores and pre-built summary tables, ask questions of it in plain English, and connect it to their own Claude with a single command. There's no trial to speak of. No plan, no paywall, no time limit, no cap on how many datasets you load. The limits that do exist are a daily spend cap on the AI queries, two ingests running at once, a 20 GB upload ceiling, and six accepted file types (CSV, TSV, TXT, Parquet, PQ, ZIP).

Q4 2026 Through the fourth quarter of 2026, Nettle completes the Silver Layer and brings its full agent fleet into production. Raw data arrives and leaves trusted: classified, cleaned, standardized and enriched automatically, with a lineage trail showing where every field came from and what changed it. Four named agents work across that foundation. Nia explores and answers questions, Neal quantifies revenue impact, Sarah owns data hygiene, and Andrew runs the data science. The work that normally takes a dedicated engineering team months now happens on upload, which means your analysts and your AI agents start from verified facts instead of guesses. Q1 2027 In the first quarter of 2027, that capability moves from proven to production scale. The agent fleet shifts from answering requests to running continuously: monitoring pipelines, catching drift as it happens, and acting before a problem reaches a report. Cross-dataset intelligence reconciles sources that disagree, showing which number is right and why. The outcome is a data operation that runs itself, with fewer people maintaining pipelines, faster answers to the questions that matter, and every decision grounded in data the business can audit.

Yes. As of September 2026, Nettle is aligning its financing documents to NVCA norms for the current round. The existing charter dates from the friends-and-family round and was adopted for speed. It will be retired once the replacement charter is complete and will not govern the new round. Gunderson Dettmer is drafting the replacement charter.

We have a policy of "your money, your vote". The members of the board are Nettle investors. The investors are informed on company progress on a regular cadence and are able to exercise their votes in an informed way. As we gain new investors, the board composition will change to reflect new capital contributions to the company.