[01]About us
Qaft builds AI systems on data other people cannot get. This is how the work is done, where the experience comes from, and what we will not take on.
[02]The sentence
The pipeline that produces clean, current, traceable data, and the agents, retrieval and interfaces that act on it.
Most AI shops only do the second half. They take a model, point it at whatever the client already has, and discover that what the client already has is stale, partial, or was never collectable in the first place. That is where pilots die, and it is not a modelling problem.
[03]Two practices
[01]
Data
Large-scale extraction, private and undocumented API integration, normalization, enrichment, freshness.
- 80 TB of public web data, refreshed monthly
- About 4M live records indexed from hundreds of sources
- Up to 100k new LinkedIn job postings indexed a day
- Four private mobile APIs unified into one typed interface at Para Inc.
[02]
Agents
MCP servers, autonomous agents, RAG over private documents, ranking and matching.
- VaraOS: financial documents into analyst-grade reports for a merchant bank
- LinkFetch: a native MCP server for Claude Desktop, Cursor and Zed
- About 4,500 jobs scored a day by an LLM, each with a plain-language reason
- Evals and a human in the loop until the output has earned the trust
A third practice only when someone else is already delivering one of the first two.
[04]How we work
[01]
No handoffs
No account manager, no delivery lead, no junior inheriting a spec. The engineer on the first call is the engineer at the handover, which is why a scope call on Monday can be running software the same week.
[02]
Agent loops, reviewed
Claude Code, Cursor and agent loops do the volume. Every compounding loop is human-reviewed until its output has proven itself. That is how a small team carries this much surface area without thinning out.
[03]
Your repo, your infrastructure
Code, prompts, models and data are yours from day one, documented as they are built. Nothing runs anywhere only we can reach, so there is no platform to be locked into and nothing to inherit if we part.
[05]Where this comes from
Qaft started in 2024. The work behind it started earlier, and this is the part of it we are free to describe.
2019 - 2022
Security and backend R&D
Reverse engineering and security research in Ankara, then B2B SaaS backends and an automated iOS and Android malware analysis sandbox, with the ingestion and reporting pipeline behind it.
2022 - 2024
Para Inc., San Francisco
A gig-economy super-app for US drivers: a serverless GCP backend at millions of daily requests, four rideshare and delivery platforms unified behind one typed interface, and the SOC 2 programme behind it.
2024 - now
Vara Capital
Qaft is the technical partner in an AI-native merchant bank and built VaraOS, the capital intelligence platform it runs on. Not a vendor relationship.
2024 - now
Our own products
UpHunt, LinkFetch, and the data platform underneath both. Built, run, marketed and supported in house, on the same stack we sell.
2026
Qaft as a studio
Taking that stack to other people's problems.
A short paid audit says whether it can be got or built, what it costs to run, and how long it takes. It also says when the answer is no.