Overview
Ecliptor is an active YC Winter 2024 company founded in 2023 by Andre Fu and Nanki Grewal. Its public product has moved through at least four frames: sales intelligence for open-source developer tools, an embedding-adapter SDK, document ingestion for complex PDFs, and Ecliptor-affiliated form-understanding research. Fu now describes his work more broadly around AI agents.[1][2]
The company has no documented shutdown or acquisition. The useful analysis is therefore a product-evolution risk audit. Ecliptor repeatedly found technically adjacent problems involving signals, embeddings, documents, and agents, but public evidence does not show which assets compounded across those changes or whether any phase became a repeatable commercial business. Current team, funding, customers, pricing, and production use remain unresolved.
Founding Story
YC identifies Andre Fu as cofounder and CEO, with prior machine-learning research, product work at Microsoft, and machine-learning work at Twitch. It describes Nanki Grewal as a University of Michigan EECS graduate focused on machine learning and systems and product design who later worked on Twitch's recommendations team.[1] YC's launch account says the founders had previously built and sold a developer tool, but it does not name the product, buyer, date, or consideration.[3]
The first Ecliptor thesis came from bottom-up adoption. Developers could use an open-source tool long before a conventional sales database recognized enterprise intent. In her February 2024 batch announcement, Grewal called Ecliptor a “new way to do bottom-up sales for OSS dev tools” and said the team had built an MVP and onboarded its first few customers.[4] Fu described the initial behavior as “auto populating your crm with developer data, directly from where they are,” contrasting person-to-person outreach with mass email campaigns.[5]
Current affiliations conflict without resolving company status. YC lists an active two-person team. Grewal's LinkedIn points to OpenAI. Fu's LinkedIn says Stealth Startup, while his personal site says, “Founder at Ecliptor, building AI agents for the future.”[2][5][6] None of these sources establishes a departure date, new ownership, or inactive company. No source explains the decisions connecting the product phases.
Timeline
- 2023: YC records Ecliptor's founding year as 2023.[1]
- February 8, 2024: Grewal announced the W24 company as a bottom-up sales product for open-source developer tools.[4]
- March 5, 2024: YC promoted Ecliptor's sales-intelligence product and its reported outreach metrics.[3]
- September 3–7, 2024: Four alpha releases of an Ecliptor Python embedding-adapter SDK appeared on PyPI.[7]
- October 21, 2024: Lantern published a jointly built private-beta compliance demonstration using Ecliptor document ingestion.[8]
- March 6, 2025: Mistral launched a competing OCR API at a stated price of 1,000 pages per dollar.[9]
- March 2026: A WACV workshop paper affiliated Fu with Ecliptor Inc. and introduced the SynthForm dataset and models.[10]
What They Built
The launch product aggregated developer interactions from LinkedIn, Hacker News, and Twitter/X to identify enterprise prospects for open-source developer-tool companies. It drafted personalized outreach from a prospect's profile and activity. Ecliptor positioned this as a bridge from a developer product's “dark funnel” to an enterprise sales process.[3]
By September 2024, the public artifact had shifted to an alpha Python SDK. Its main operation accepted an embedding and returned that embedding multiplied by a customer-specific adapter. PyPI shows four releases over five days, names Grewal as author and verified maintainer, and shows no later release in the observed record.[7] No source explains how the adapter related commercially to sales intelligence.
The October product was document infrastructure. A Lantern tutorial called Ecliptor's PDF ingestion and smart-chunking endpoints, converted banking-compliance PDFs into Markdown, preserved tables and structure, and stored embeddings in Lantern for retrieval and LLM-based checks. Smart chunking could bring relevant context from elsewhere in a document into a semantically useful unit.[8] The companies described Ecliptor as a private beta for financial-services teams with complex documents.
That tutorial is functional evidence, not a customer case. Lantern and Ecliptor jointly built it using public regulations and synthetic customer-support interactions. It does not establish a production bank deployment.[8] The live Ecliptor demo still converts PDFs to Markdown chunks for embedding and retrieval, with attention to tables, images, and chunk boundaries, while the root site is only a “Coming Soon” page.[11][12]
The 2026 SynthForm work moved closer to model research. It introduced 3,417 synthetic samples across tax, immigration, finance, healthcare, dental, and insurance forms, plus 2B, 4B, and 8B models. The authors report 66.0%, 69.6%, and 70.5% ANLS on their own test set and publish the dataset, checkpoints, and training code.[10] This is recent inspectable Ecliptor-affiliated work, not proof of a current commercial model.
Market Position
Target Customers
The buyer changed with the product. Sales intelligence addressed open-source developer-tool vendors. The adapter SDK addressed embedding users. Document ingestion targeted financial-services teams handling complex PDFs. SynthForm focuses on degraded scans and faxes in high-stakes workflows where errors in birthdates, medical codes, Social Security numbers, or policy IDs can break downstream processes.[10] Fu's AI-agent framing broadens the direction again without identifying a buyer.
Market Size
No observed source provides a defensible serviceable market for Ecliptor's current product. Broad sales-tech, document-AI, and agent categories would create false precision because the current commercial scope is undocumented. No customer count, page volume, paid-seat count, or deployment footprint is available.
Competition
Common Room overlaps the original product by resolving buying signals across social, community, product, code-collaboration, and CRM channels to people and accounts.[13] In document ingestion, Reducto parses text, tables, figures, locations, confidence, and chunks for LLM workflows.[14] LlamaIndex's hosted parser returns Markdown or text with optional bounding boxes.[15] Unstructured partitions messy files into structured elements for retrieval and transformation.[16]
Mistral OCR adds model-level price pressure by extracting ordered text, images, tables, equations, and other document structure, with public pricing and selectively available self-hosting.[9] A private-beta parser with no public pricing, security terms, or production proof must beat both specialized vendors and model vendors on a narrower workflow. Ecliptor's public record does not establish that differentiation.
Business Model
No public pricing, revenue, contract structure, retention, gross margin, document volume, burn, or runway is available. The sales product likely used business-to-business contracts; the document demo routes interested users to Fu, suggesting a sales-assisted motion. Neither observation proves packaging or paid use.
Funding evidence is internally inconsistent. Dealroom identifies YC as an investor but shows $125,000 in one field and $500,000 in narrative text. It shows no later visible round, yet the page is incomplete and cannot prove that no other capital was raised.[17] No primary financing announcement resolves the amount.
The commercial risk across phases is loss of compounding economics. Sales-intelligence data, embedding adapters, PDF parsers, form models, and agent products can each support a company, but they have different buyers, distribution, costs, and durable assets. Public evidence does not show that one phase financed or strengthened the next.
Traction
YC's March 2024 launch post reported 100% email opens, about 25% replies, and about 10% calls booked among existing customers.[3] These are promotional figures without sample size, campaign design, attribution method, or independent audit. Grewal's “first few customers” establishes early adoption, not retention or scale.[4]
Later evidence shows technical activity rather than commercial traction: four alpha SDK releases, a synthetic partner demo, a live ingestion demonstration, and a published research artifact. No named production customer, customer-authored case study, current pricing, or revenue metric was found. The Lantern collaboration should not be promoted into a banking-customer deployment.
Post-Mortem
Ecliptor remains active in YC's directory, and recent research uses its affiliation. This section audits product and operating risks rather than inventing a terminal event.
Technical adjacency can hide commercial discontinuity
Signals, embeddings, chunks, forms, and agents share machine-learning components. Customers do not buy that adjacency. Developer-sales teams, retrieval engineers, financial document operators, and form-processing owners have different budgets and acquisition channels. Without a documented carryover of customer data, distribution, or workflow ownership, each pivot may restart commercial learning even when it reuses technical expertise.
The counterargument is that a young team should explore quickly. The short SDK release burst and Lantern demonstration may have been experiments on a path toward form intelligence, not discarded companies. The evidence does not reveal whether this was disciplined search or repeated reset.
Parsing became a priced commodity
Reducto, LlamaIndex, Unstructured, and Mistral compete on document structure and downstream readiness. Mistral's 2025 price of 1,000 pages per dollar makes generic OCR and Markdown conversion difficult to defend on ingestion alone.[9] Ecliptor's deeper form research may answer that pressure, but author-reported results on a synthetic dataset do not establish accuracy on a customer's changing legacy forms.
Public technical work outruns commercial proof
The live demo and 2026 paper show continued capability. The placeholder root site, absent pricing and security material, no named production users, and conflicting founder affiliations leave the commercial system unclear. That is an execution and disclosure risk, not evidence of shutdown.
The exact founder quotes available describe the original sales product and Fu's current agent framing; none explains the transitions or a failure cause. A post-mortem quotation would therefore be fabricated. The strongest conclusion is bounded: Ecliptor has preserved technical motion while its repeatable business and current operating team remain unverified.
Key Lessons
- Technical adjacency is not customer continuity. Shared ML components do not guarantee that buyers, distribution, or proprietary assets carry across pivots.
- Promotional conversion needs a denominator. Eye-catching percentages cannot establish repeatability without sample size, attribution, and retention.
- A demo is not deployment. Synthetic partner work can prove integration while saying nothing about production adoption.
- Research can answer commoditization only with field proof. Better form models matter commercially when they survive real template drift and downstream dependencies.
Sources
- Y Combinator, Ecliptor company profile.
- Andre Fu, personal site.
- Y Combinator, Ecliptor launch post, March 5, 2024.
- Nanki Grewal, W24 announcement, February 8, 2024.
- Andre Fu, LinkedIn profile.
- Nanki Grewal, LinkedIn profile.
- PyPI, Ecliptor SDK.
- Lantern, Ecliptor compliance tutorial, October 21, 2024.
- Mistral, OCR launch, March 6, 2025.
- WACV 2026 SynthForm paper.
- Ecliptor document-ingestion demo.
- Ecliptor root site.
- Common Room Signals.
- Reducto Parse documentation.
- LlamaIndex parsing documentation.
- Unstructured API documentation.
- Dealroom, Ecliptor company profile.