16 partial matches in Q4 2025 — search/RAG and coding-agent startups absorbed, but almost none pass the origin criteria
Asked:“expand headcount to 200-300 max now but was smaller in 2025. include Cypurs or other locations given key tech experts from Russia, Belarus. more focus on casual and agentic vs graph rag. ok if company is not closed but aquired and rag product is discontinued. report more partial matches and raw data”
The expanded screen covers October–December 2025 and surfaces 16 partial matches across five segments: enterprise search/RAG (4), coding/DevOps/security agents (6), general or workflow agents (3), enterprise agents/knowledge (2) and AI infrastructure with a discontinued external product (1). These are partial matches, not strict hits: the crawl ran English and Russian queries, yet verified Russia/Belarus technical links were sparse, and no solid Cyprus, Belgrade or Kazakhstan presence was verified for the final shortlist.
Strictness ladder — how many of 16 pass each rung
Location/origin rungs are nearly empty: no verified Russia/Belarus technical link and no verified Cyprus/Belgrade/Kazakhstan presence in the shortlist. A Russian-registered “Gradient Labs” appeared in raw extraction and is excluded — the user excludes Russia-based companies, and it is unrelated to DeepSky (formerly Gradient, acquired by Airtable). Never infer origin from a Slavic name; Neptune.ai’s founders are Polish.
Q4 2025 timeline by segment
enterprise search / RAG
agentic / coding
other enterprise / infra
acquired (circle)
product discontinued (square)
◆ chronology flag — verify manually
Confidence: ● primary/official source◐ corroborated secondary○ single-source lead — month-only dates are plotted mid-month and marked “~”.
Findings
Strongest discontinued/absorbed-product cases:
Neptune.ai (OpenAI acquisition announced 2025-12-03; the hosted external service wound down after transition — Warsaw roots, 51–200/73 employee estimates, Samsung/Roche/HP among 60,000+ users, >$18m raised) —
7wdata.beJina AI is the cleanest primary-source case: its acquisition by Elastic completed 2025-10-09, with embeddings, rerankers, readers and small language models for multimodal/multilingual search —
times-online.com
Codegen (ClickUp, Dec 2025; standalone version later unavailable) and
Codeium/Windsurf (Cognition transaction; 210 people, 350 enterprise customers retrieved — chronology noisy across July and December 2025 steps, do not flatten to one closing date) —
aiodyssey.io,
chatforest.com
Manus AI (~100 people, Meta’s ~$2bn deal announced 2025-12-29) was later
unwound by Beijing’s NDRC and repurchased via Tencent, HongShan and ZhenFund — not a permanent acquisition —
dlvr.it
Sana Labs (Workday, $1.1bn) is over the preferred headcount cap (~300 or ~592 in retrieved sources) and ranks low as a context match;
Tromzo appears very small (~7 in a later source);
DeepSky had ~12 team members and $40m referenced —
agentic.ai,
chroniclejournal.com,
7wdata.be
Shortlist — all 16 partial matches
Enriched raw data — 12 entities
Blanks preserved as “—”. Dates may reflect announcement vs completion; both are retained where sources conflict.
Appendix — machine-extracted people/location signals (noisy)
Uncleaned search extraction. Clean signals: Windsurf 101–250 staff; Sana ~300 or ~592 (over cap); Neptune.ai 51–200 / 73 and “Polish”; Tromzo ~7 (later source); Codeium ~40 employees hired in the July 2025 transaction; Manus in Singapore/China; General Agents in San Francisco/worldwide. Russia/Belarus fields are mostly null. The Russian-registered Gradient Labs row is shown struck through: excluded.
Method: expanded follow-up to the Q4 2025 screen; qualification-first, audit-friendly. Data: a 16-row Q4 2025 shortlist (segment, event, sources), 12 enriched company rows, and a 10-row noisy machine extraction of people/location/headcount signals, retrieved via English and Russian queries in early 2026. Inclusion broadened to ≤~300 current headcount (smaller in 2025 preferred) and to acquisitions where the AI/RAG/agent product stopped being standalone. Mentions = count of retrieved sources naming the event. Event dates mix announcement and completion; conflicts are flagged, not resolved. Verbose extraction text truncated for space; full source URLs linked per row.