POMA AI
Tomorrow's agents cannot use yesterday's search.
Document ingestion is an established, paid market: Azure Document Ingestion, AWS Textract, Unstructured, LlamaParse, Reducto, Mistral OCR. They charge a few cents a page and hand you text. Your agents will choke on that.
POMA hands you retrieval-ready chunks — the same ingestion, done better, with the structure-aware chunking that decides whether retrieval works. Now that agents, not humans, read the results, that quality is the whole game.
Rolling CLA · €8M cap · 25% discount · €500K remaining
Search's reader changed — from human to agent.
For decades recall was king — because humans consumed the results, forgiving readers who skim a list, skip the junk, and stitch the useful pieces together. Agents don't read like us: they pull results straight into a limited context window, attend to everything at once, and a single plausible-but-wrong passage derails the task with no amount of reasoning able to recover. Retrieval's job just flipped — from "return everything relevant" to "admit only clean, correct context."
The Human Era — Tolerant
Skims, skips the junk, stitches the pieces. Recall was king: return everything plausibly relevant — the human sorts it out and forgives the noise.
Needs forgiveness.
The Agent Era — Fragile
Ingests it all into context, attends in parallel — can't skip. One plausible-but-wrong passage derails the whole task — and no amount of reasoning recovers.
Needs precision.
Search is ancient. What POMA solves is brand new.
Three technologies converged in the last few years — dense vector retrieval, LLMs that can read implicit document structure, and agents that consume retrieval directly. POMA sits exactly where they meet.
1960s–2010s — The keyword era
Inverted indexes, BM25, PageRank. Fifty years of incremental refinement on the same core idea: match terms, rank by frequency.
2013 — word2vec
Words become vectors. Semantic similarity becomes computable for the first time.
2022 — Vector search at scale
Dense retrieval goes production-ready. Meaning, not keywords, drives what gets returned.
2023 — LLMs read structure
Models can now infer hierarchy, section boundaries, and implicit document architecture — not just surface text.
2025 — Agents do the searching
Retrieval is no longer a human-facing UI. Agents pull context directly, which goes straight into reasoning.
Now — POMA
Structure-aware ingestion that emits retrieval-ready chunks. Built for the era where agents do the reading.
Every retrieval pipeline starts with something broken nobody talks about: the chunks.
A chunk is the atomic unit of retrieval — the exact slice of text your vector database stores, searches, and returns to the agent.
Every ingestion vendor — Azure DI, AWS Textract, Unstructured, LlamaParse — parses your documents to text, then stops. They hand you a wall of markdown and leave you to split it into chunks yourself. Usually with a free, naive splitter that knows nothing about your document's structure. That broken chunk is what your vector database indexes. That broken chunk is what retrieval runs on. That broken chunk is what your agent reasons over.
Naive splits
500-token blocks that start mid-sentence and end mid-table
Lost context
No section, no hierarchy, no position — just floating text
Dirty text
OCR artifacts, page headers, watermarks embedded as content
You can't out-context a bad chunk.
You can't fix a bad chunk downstream at retrieval. The damage is done at ingestion. Agents attend to every retrieved token at once, so noise isn't ignored — a distractor actively degrades the answer. Accuracy drops as input grows, with facts buried mid-context retrieved worst. —> Context Rot.
Sources: ChromaDB context-rot study and related academic work (Lost in the Middle, NoLiMa, RULER)
POMA PrimeCut
Better document ingestion
— with chunking built in
Every ingestion vendor hands you text
and leaves the chunking to you.
POMA PrimeCut does both —
structure-aware ingestion that emits retrieval-ready chunks, not raw markdown.
Same price point as OCR. Fundamentally better output.
Same document. Three very different chunks.
Here's the same passage from an FDA cybersecurity guidance document, chunked three different ways.
500-token naive split
...an SPDF is one approach to help ensure that the QS regulation is met. Because of its benefits in helping comply with the QS regulation and cybersecurity, FDA encourages manufacturers to use an SPDF, but other approaches might also satisfy the QS regulation.
### B. Designing for Security
When reviewing premarket submissions, FDA intends to assess device cybersecurity based on a number of factors, including, but not limited to, the device's ability to provide and implement the security objectives below throughout the device architecture.
Security Objectives: • Authenticity, which includes integrity • Authorization • Availability • Confidentiality • Secure and timely updatability and patchability
...The risks presented by cybersecurity vulnerabilities; the exploitability of the vulnerabilities; and the risk of patient harm due to vulnerability exploitation.
### C. Transparency
A lack of cybersecurity information, such as information necessary to integrate the device into the use environment...[truncated — chunk continues across §C]
- Spans multiple sections
- Heading isolated from its content
- Bleeds into unrelated topic (Transparency)
Incumbent parser
• The device's intended use, indications for use, and reasonably foreseeable misuse;
• The presence and functionality of its electronic data interfaces;
• Its intended and actual environment of use;18
• The risks presented by cybersecurity vulnerabilities;
• The exploitability of the vulnerabilities; and
• The risk of patient harm due to vulnerability exploitation.
- No section indication
- No position within document
- OCR artifact in main text ("18")
Hierarchically prefixed chunk
Cybersecurity Guidance for Medical Devices
└ Guidance for Industry and FDA Staff
└ B. Designing for Security
└ The extent to which security requirements, architecture, supply chain, and implementation are needed to meet these objectives will depend on but may not be limited to:
└ Its intended and actual environment of use:
└ The risk of patient harm due to vulnerability exploitation.
What makes it so much better:
- Full document path preserved
- Self-contained meaning
- Zero artifacts, zero ambiguity
Tokens at 100% Recall — the key metric for the agentic era, and POMA wins it.
On the public POMA-OfficeQA benchmark, PrimeCut achieves 100% recall using just 23% of the context tokens that naive pipelines need — 4× less noise, cost, and context rot per query.
Databricks + naive chunking
500/100 split/overlay strategy
Unstructured.io
Default "hierarchical" chunking
POMA PrimeCut
Full recall, 4× less context
Public and reproducible: github.com/poma-ai/poma-officeqa (based on the hitherto unsolved Databricks OfficeQA challenge)
Embedders love structure.
Content hierarchy serves as a coordinate system
Dilution comes from incoherence, not length
Structural markers are semantic anchors
Fixed-size chunking feeds embedders what they hate
rank 1,780 → 15 (119× more relevance)
Ranking improvement, same model & cosine, with only the chunking changed (conventional → POMA)
153× improvement
by prefixing the full hierarchical "content path" to 21,414 exemplary chunks
PrimeCut: Ingestion plus best-in-class Chunking
PrimeCut reads document structure — sections, tables, lists, figures — and emits hierarchical, retrieval-ready chunks instead of arbitrary slices. One product replaces your OCR/parser and the chunking step the others leave to you.
50+ filetypes
PDFs, tables, figures, code, and more, including content most parsers flatten
Hierarchical chunksets
Patented, structure-aware; no meaning lost at boundaries
Shipped and proven
Live API, public benchmark, US patent granted
From €0.003/page
At the OCR price floor, chunking included
An established market, entered with a better product
Ingestion is already a budget line — trillions of pages parsed per year at a few cents each. We don't need to create demand; we need to win switchers. Mistral OCR proved in 2025 that cloud buyers will move to a new entrant on price and quality.
Proven category
Azure, Textract, Unstructured, LlamaParse, Reducto, Mistral OCR all sell it today.
Clear wedge
They output text; we output retrieval-ready hierarchical chunks, at or below their price*.
Market focus
Many industries process complex documents — we focus on strong signals from their go-to integrators.
*PrimeCut Eco sits at the market floor (0.3¢); PrimeCut Pro (3¢) undercuts most structured tiers.
Integrators don’t switch — they constantly onboard.
Agencies, consultancies, and RAG platform builders are the ideal first market. They don’t need to migrate existing pipelines — every new client they win lands on POMA from day one.
Early channel partners
Integrators like ki-ra and Heuristiq are already building on POMA. Each new project they take on is a new POMA customer.
Public benchmark as driver
Public and reproducible (github: poma-officeqa). Gives integrators the proof they need to sell the switch to their clients.
The retrieval pros are not competing — they are waiting for POMA:
"If only our customers had better chunks…"
( principal engineers from both vespa.ai and hornet.dev at the Applied AI Conference 2026 )
Vector databases start at the chunks. Their own docs tell you to "split text into chunks yourself."
PrimeCut produces exactly what is needed by Vespa, Hornet, Turbopuffer, qdrant, Pinecone, Weaviate, or any other engine.
Built to defend 75%+ margin
Patented core chunking IP
Structure-aware ingestion and hierarchical chunking in one pipeline, not replicable from off-the-shelf parts.
Public benchmark proof
100% recall at 23% of the context tokens naive pipelines need — 4× less noise, cost, and context rot.
First agent-ready context format
Agent pipelines will need better chunks. Customers who've built on POMA's chunksets won't go back.
Usage-based model
Pay per page, free tier to start, enterprise on request. Grows automatically as customers process more.
Experienced team, already in place
Leadership
Dr. Alexander Kihm — Founder, CEO & MD
PhD, Big Data Econometrics (German Aerospace Center).
Serial entrepreneur (Advo Assist, fairr — exit to Raisin).
25+ yrs coding.
COO & future Co-MD — Dr. Jan Simon Raue
Confidential, under NDA. Joining September 2026.
Serial entrepreneur, PhD, 20+ yrs building startups; takes over operations and commercial scale.
Product & Engineering
- Ryan — Director of Product, 15+ yrs
- Fabia — Senior Core Developer, PhD, 10+ yrs
- Raffael — Senior DevOps, 10+ yrs
- Sepehr — AI/ML Researcher, 10+ yrs
- Alex — Solutions Engineer, BSc on RAG
Business & Sales
- Annika — Head of Marketing, 10+ yrs
- Florian — Business Development, 7+ yrs
- Jens — CFO & Legal, external/part-time, 20+ yrs
A rolling CLA at constant terms
Rather than orchestrate a formal round, we've signed CLAs continuously at the same terms for the past year. The €8M cap hasn't moved while the company has — patent granted, Grill shipped, design partners live. The remaining 500k of investments enter today on the same terms as those last year.
CLA
€8M cap · 25% discount · €500K remaining of the €1.5M target
Public grants & loans
~€900K, expected Q3 2026. In process.
Cloud credits
>$200K (Google Cloud, AWS, Azure, Stripe). Runway through 2026.
Use of funds: Product & Engineering 55% · Go-to-market 30% · Operations 15%.
Investors (extract): heartfelt.vc · B# (b-sharp) · Till Faida (founder, Adblock) · Philipp Rechberg (Caya) · Steffen Reitz (gini)…
Runway to late 2027 incl. public funding.
Next steps & timeline
H2 2025
Company founded · Primecut (ingestion + chunking) live, 50+ formats · first design partners ✓
Q1 2026
Public benchmark · US patent granted · first revenue ✓
Now
POMA Grill · COO & Co-MD signed (joins Sep) · opening self-serve
H2 2026
Developer-led growth · first paid expansion · scale usage
2027
Sales-assisted motion at the top · priced seed round
Ground-floor investment in the agentic AI future
Search's quality bar has moved for the first time in decades — from recall to clean, complete context — because agents now do the reading. POMA is built for that shift: ingestion that ships the retrieval-ready chunks everyone else makes you build yourself.
A patented core, a public benchmark, and the path into one of software's largest horizontal markets.
Contact: Dr. Alexander Kihm · ak@poma-ai.com