Rossen Genchev
01 / 18Software veteran · medals and scars earned in the trenches, not at headquarters

Rossen Genchev

Being early looks exactly like being wrong, for about four years. I have been early thirteen times, from search engines in 1996 to LLM agents in 2020. Below: the thirteen, dated against the day the industry caught up, and what each one became.

linkedin.com/in/rossengenchev[email protected]

02 / 18Two receipts

Before the argument, the receipts.

$1B
Dynamo Software, valued in 2021.
Eleven years there, 2002–2013, three of them as Chief Software Architect.
€1M
Practicedent, sold to CURADEN AG.
First company, built and sold within two years.
03 / 18The claim

Being early is indistinguishable from being wrong.

For about four years. Then the thing gets a name and everyone agrees it was obvious. I have lived inside that gap thirteen times. The receipts are above; the rest is measured below.

04 / 18What I am for

Four things.

  1. 01System design and software architecture
  2. 02Technical leadership and managing engineering teams
  3. 03Distributed software systems
  4. 04LLM agent integrations
05 / 18The refusal

Ignore this CV if

  1. 01the phrase you hear most in this building is “that’s hard”, not “that’s interesting”;
  2. 02people here grow by who they had lunch with, not by what they shipped;
  3. 03you are still waiting for AI to prove itself;
  4. 04“let’s wait and see” has ever won an argument here;
  5. 05you need someone else to have named a thing before you will fund it.

Everyone else, read on.

06 / 18Plate I · what and when

Thirteen systems, from built to named.

The gap is closing. Not because I slowed down: the industry got faster.

  1. 01Inverted index, compressed posting lists, proximity rankingon a 386 with 4 MB, Windows 3.11996 → Lucene, 1999
  2. 02Cooperative concurrency: every query a state machine stepped on idleno OS threads available1997 → State Threads, 1999
  3. 03Radio and TV archives digitised to MP3, a portable player reaching themBrandCom, later acquired by a US TV network2000 → podcasting, 2004
  4. 04Unit tests, derived from watching QA repeat what code could checknot read anywhere2001 → TDD, 2002
  5. 05HTML as the user interface of a desktop applicationfirst pilot, WebMessenger2002 → Electron, 2013
  6. 06Attribute-annotated objects bound to database operations and UI controlsDynamo, .NET2003 → Entity Framework, 2008
  7. 07UI generated from the schema; self-arranging layout reshaped by drag and dropa grid renderer for WinForms2004 → responsive design, 2010
  8. 08A backend made only of plugins with manifests and a global event busDynamo 62011 → event-driven microservices, 2014
  9. 09EDL: one declarative language for schema, process and screensthe foundation of Dynamo 62011 → low-code platforms, 2018
  10. 10Shopping by sentence, not by keywordHashStyle, on the language models of 20152015 → LLM shopping assistants, 2023
  11. 11Bet the company on language models, days after GPT-3the client I wrote to was skeptical2020 → ChatGPT, 2022
  12. 12The model reads, generated code verifies, the mismatch goes back as a promptfinancial tables with merged cells~2022 → verifier loops, 2024
  13. 13An LLM-facing layer in front of every data source; forms assembled inside the chatprototype rejected as too unfamiliar2023 → MCP, 2024
07 / 18Chapter I · 1994 – 1998

The self-taught

Specialised school for electronics and computing · Sofia University, BSc Mathematics, 1995–1998 · Ciela, senior programmer, 1996–1997
  1. 01At 17, learned C++ from a Russian-language book without knowing Russian. Wrote a DOS program that plots any typed function: expression translated to reverse Polish notation, numeric roots, differentiation, integration, a windowed interface. No internet.
  2. 02At 18, hired straight in as senior programmer at Ciela, a legal information system. Requirement: full-text search on a 386 with 4 MB of RAM, under Windows 3.1, serving several searches at once on an OS without threads.
  3. 03Built it in six months: dictionary stemming, noise removal, inverted index, posting lists sorted and compressed in blocks of 10,000, merge-style intersection that never fit in memory, proximity scoring on the survivors. Concurrency by hand: each search a state machine advanced one step per idle message.
08 / 18Chapter II · 1999 – 2001

The reluctant commander

BrandCom, Israeli startup · programmer, then CTO at 21
  1. 01First web work. Built pizza911, one of the first online shops in New York, in C++ CGI.
  2. 02Digitised and archived radio and television audio using MP3 while it was still an exotic format; prototyped a portable player that reached the archives. The archiving product succeeded; the company was later acquired by an American TV network.
  3. 03While the founder was abroad, the team was stalling. Brought in process and discipline with no title; people followed. On his return he made me CTO. Ran about 25 people across 10 projects for two years, then left: the pressure was more than I wanted to carry.
Extinct branch: the portable archive player. Three years later the world called the idea podcasting.
09 / 18Chapter III · 2001 – 2002

The heretic

WebMessenger, senior programmer · Microsoft .NET Beta 1 programme
  1. 01Introduced unit testing to the company. I had not read about it; I had watched QA spend days on checks a program could run in seconds.
  2. 02Ran the first pilot using HTML as the front end of a desktop application, for effects that were painful in native controls.
  3. 03Member of Microsoft’s .NET Beta 1 group a year and a half before release, feeding back on C#/COM interop and on memory owned by unmanaged code. Argued that C# or Java should replace C++ for business logic. Management declined. I left.
10 / 18Chapter IV · 2002 – 2013

The architect

NetAge Solutions, later Dynamo Software · senior developer, team lead, then Chief Software Architect for three years
  1. 01Joined to pilot C#. Delivered TTFSE, an EU-sponsored logistics, route-planning and customs-documents system, on time: first web services, first interactive route map.
  2. 02Moved to Dynamo, a CRM for investment bankers that lived entirely inside Microsoft Exchange, with Outlook as the client and the Exchange store as a document database. Filtering, aggregation and relations were thousands of lines of if/else. Replaced them with an XML metadata layer: schema changes went from slow and fragile to routine.
  3. 03For the full .NET rewrite, built an attribute-based binding engine mapping annotated objects and controls to database operations, then removed hand-built WinForms screens altogether: the UI is generated from the business schema, laid out by a responsive engine with an HTML-table model inside, and reshaped by users with drag and drop. The product became more flexible than its competitors. That work earned the title.
  4. 04For Dynamo 6 (2013): designed EDL, the Entity Definition Language, covering business-object-to-schema transformation, process descriptions and control descriptions, and a backend built entirely from plugins with manifests and a global event bus.
  5. 05Left in 2013 to found my first company. Dynamo went on to become a category leader: Francisco Partners majority stake in 2017, Blackstone growth investment in 2021 at a reported valuation approaching $1B, over 1,000 clients today. Dynamo is now a BioSeek client.
11 / 18Chapter V · 2013 – 2018

The founder, twice

MediCloud, co-founder, 2013–2015 · HashStyle, co-founder, 2015–2018
  1. 01MediCloud started from a personal collision with healthcare: a patient-owned health record, readable and interactive rather than hundreds of pages. Too large for the team. Pivoted to dental practices, where we had an investor.
  2. 02Built Practicedent, a cloud CRM/ERP for dental clinics in Switzerland and Germany, in more than ten languages: a “mini Dynamo” on the MEAN stack while every part of it was brand new, with JSON flowing unchanged from browser to database. About 90% less code than the original. Acquired by our investor CURADEN AG in 2015; the founders exited for around €1M after two years.
  3. 03HashStyle: an e-commerce platform that understood sentences, not keywords, on the primitive language models of 2015. Came within reach of an acquisition by Viber. The deal fell through and the company nearly went under.
Extinct branches: the patient-owned record, and HashStyle itself. Descendants: the one-language, one-representation stack I have used since, and the semantic search that became BioSeek.
12 / 18Chapter VI · 2018 – today

The cartographer

BioSeek JSC, CEO and co-founder · bioseek.eu, aerosaas.ai
  1. 01A biotech client asked for a better search engine. What the text actually contained was a knowledge graph waiting to be extracted. BioSeek was founded to build it.
  2. 02Wrote an entity-recognition algorithm that finds around one million genes, compounds, organisms and drugs in a page of text in under a millisecond, with statistics and careful memory layout rather than machine learning. Licensed to Dynamo Software, where it links customer databases to document corpora for 42 clients, including Lockheed Martin and the University of Pennsylvania.
  3. 03Moved the company to language models in 2020, days after GPT-3. Receipt recognition for a fuel company at 99% accuracy with OCR plus GPT-3.5, before vision models existed. A CV platform for a recruiter: PDFs to strict JSON under a fixed ontology, semantic search with reranking, and matching that splits each CV and job ad into semantic cores, scores each with the model, and lets the model run the conversation until the gaps are filled.
  4. 04For Dynamo: natural language translated into their filtering meta-language, saving experienced users 10–15 minutes a query, with layer-by-layer schema extraction because the context window could not hold the schema; extraction of financial data from Excel and PDF with generated Python checking the arithmetic and feeding mismatches back; and a prototype where every data source sits behind an LLM-facing layer and every operation runs from chat, with forms assembled on the fly. Rejected as too unfamiliar, a year before MCP.
  5. 05A phone-line pipeline (noise removal, specialised speech-to-text, an adapted model correcting transcription errors) reaching near-perfect success at 5–8 seconds of latency; a US legal knowledge graph across federal and state law, amendments and court audio with speaker identification, queried by agents forbidden to use their own legal memory.
  6. 06AERO: agents interview a client and build the complete operating system of the business, CRM to ERP to documents. Not by generating and testing millions of lines of code, but by writing BODL, the Business Object Definition Language, EDL’s direct successor, which an engine executes.
  7. 07For an e-commerce company (under NDA): an agentic commerce platform. Sixteen coordinated agents take a shop from a structured interview to a running business, from migration to merchandising, retention and a daily briefing to the merchant; the store engine is built for Shopify-class load and scale. Behind it a back-office ERP, written in BODL and run by AERO, close to a full counterpart of the large enterprise suites.
Paused branch: the full public-biology graph. Converting hundreds of millions of pages costs millions in inference even at the cheapest models. The engine exists; the fuel does not, yet.
13 / 18One idea, six times

The same move, one level higher each time.

Every period of my career is the same move, one level higher: describe the thing, then let an engine execute the description. The tool in my hands evolved. The idea did not.

  1. 1995RPN interpretera function typed, then plotted
  2. 1996Inverted indexdocuments described as term sets
  3. 2003XML metadata layerfilters and relations without code
  4. 2011EDLschema, process and screens in one language
  5. 2018Knowledge grapha domain described as entities and relations
  6. 2025BODLagents describe a business, an engine runs it
14 / 18Everything, in order

Twenty-four systems.

A small share of what I have shipped: the ones kept here are the ones where the hard part was the engineering.

  1. 1994DOS math function plotterC++
  2. 1996Ciela full-text searchC++, Delphi, MFC
  3. 1999pizza911 online shopC++ CGI
  4. 2000Radio/TV MP3 archiving, portable playerC, C++
  5. 2001WebMessengerC++, MFC, DirectX, ATL
  6. 2002TTFSE logistics and customsC++, VB, C#/.NET, MS SQL, HTML/JS
  7. 2003Dynamo on Exchange, XML metadata layerC/C++, COM, VB
  8. 2003Dynamo .NET binding engineC#, COM, C++
  9. 2004Dynamo schema-generated UI and layout engineC#
  10. 2011Dynamo 6: EDL and plugin backendC#, C++
  11. 2013MediCloud / PracticedentMEAN
  12. 2015HashStyle semantic e-commerce search.NET
  13. 2018BioSeek entity-recognition algorithm.NET, Go
  14. 2018BioSeek biotech knowledge graph.NET, JS, Go, Neo4j, MongoDB
  15. 2021Dizelor receipt recognitionMEAN, LLM
  16. 2022Arena CV ingestion platformMEAN, LLM
  17. 2022Arena candidate–job matchingMEAN, LLM
  18. 2022Dynamo text-to-gridC#, Go, Python, LangChain
  19. 2022Dynamo Excel/PDF extractionPython, LangChain, LangGraph
  20. 2023Dynamo LLM-native layer and chat frontendPython, LangChain, LangGraph
  21. 2024LLM on a phone linePython, LangGraph, MEAN, Elasticsearch
  22. 2024US legal knowledge graphPython, LangGraph, PostgreSQL, D3.js
  23. 2025AERO / BODLPython, LangGraph, MEAN, React, D3.js
  24. 2026Agentic e-commerce platform and back-office ERP (under NDA)Python, LangGraph, BODL / AERO, MEAN
15 / 18Education
  1. 1995–1998Sofia UniversityBSc Mathematics
  2. to 1995Specialised secondary school for electronics and computing machinesBulgaria

Publications

  1. MediumDreaming in Code: Envisioning the Perfect Programming Language (DreamLang)original
  2. 2026Metaprogramming: An elegant weapon for a more civilized ageoriginal
  3. 2026The AI Apocalypse Won’t Come From Superintelligence — It Will Come From Layoffsoriginal
  4. From Neuron to Agent: What We Learned When We Tried to Speak to Machinesoriginal
  5. 2026Mathematics Is the Most Beautiful Thing in the World. School Made Sure You’d Never Find Out.original
  6. Erlang: The Curiously Strong Language Every Developer Should Learnoriginal
  7. Redis: From Simple Key-Value Store to One-Stop Database Solutionoriginal
  8. Metamorphosis to Greatness: How Startups Transformed Through Pivotingoriginal
  9. The Tyranny of Bureaucracy and the Heroic Individualoriginal
  10. 2023Unearthing the Hidden Genius of ChatGPT and Its Ilk: A Lobster’s Guideoriginal
  11. 2023Adopt GPT, or Be History!original
  12. 2023The Hitchhiker’s Guide to Artificial Intelligenceoriginal
  13. 2023Leadership Lessons for Business from The Art of Waroriginal
  14. 2023The Quantum Computing Tsunami: A Looming Threat to Cryptocurrency Securityoriginal

Five books that shaped the work

  1. 01The Nature of Software DevelopmentRon Jeffries
  2. 02Enterprise Integration PatternsHohpe and Woolf
  3. 03Lean from the TrenchesHenrik Kniberg
  4. 04Seven Concurrency Models in Seven WeeksPaul Butcher
  5. 05Learn You Some Erlang for Great Good!Fred Hébert
16 / 18Where to look

See it working.

Two products live today, and four recordings of systems from the index, running. Each recording has its own page, with what it shows and a transcript.

bioseek.eubioseek.euBioSeek, the biotechnology re:search engine.The questions a drug programme turns on — which targets hold up, what is patented and by whom, where the safety signals are, who else is in the race — as one graph the agents reason across, answered with the sources attached.aerosaas.aiaerosaas.aiAERO, agents that build the operating system of a business.Agents interview a client and write BODL, the Business Object Definition Language; an engine executes it. CRM to ERP to documents, without generating a line of code.
  1. 2019BioSeek, the biotechnology re:search engineOne minute from 2019: the questions a keyword search cannot answer about a disease, and a search engine that answers them with the sources attached.YouTube · 1:03 · Transcript
  2. 2025Contract analysis: clause risk, scored and sourcedA service agreement uploaded, read and split into clauses; each clause analysed for risk, given revised wording, and linked to the case law and regulation it rests on.Loom · 3:02 · Transcript
  3. 2025Public procurement: irregularities and live pricesA public tender read by agents: the specifications quietly written around one supplier flagged, and every line item priced against market alternatives the system finds and cites.Loom · 4:39
  4. 2017Semantica: search for fashion, by sentence (HashStyle)A shop searched by asking in plain language, on the language models of 2015. The engine underneath became BioSeek.Vimeo · 1:44
17 / 18Colophon

How this was made

  1. 01Fifteen minutes to set up, one instruction: a CV that does not blend into the pile. The agent worked the roles in turn — designer, copywriter, illustrator, critic — and built its own instruments as it went: a check that fails when a page overflows A4, a comparison that shoots every page from the browser and from the PDF and reports what differs, a filter that flattens each drawing onto exactly three inks.
  2. 02The illustrations are generated. The charts are d3. No design tool was opened.

Item 04, demonstrated rather than asserted.

18 / 18Contact me if

Being early has a price, and I have paid it more than once: a language proposal declined, a chat-first interface declined, a company left before the exit, a deal that did not close. I would rather keep paying it than arrive on time with everyone else.

You are building something that does not have a name yet.

Download the CV · PDF, 9 pages[email protected]
Rossen Genchev, September 2026. [email protected]linkedin.com/in/rossengenchev · bioseek.eu · aerosaas.ai