Consulting
Tell us your scope: book a 20-minute call.
Deemwar is an implementation partner for software teams and startups: we build and ship alongside your engineers, take overflow work for dev shops, make your product and team AI-native, and handle cloud, data and migration work of any shape.
Book a 20-minute callBuild and ship with your team
Features, services and fixes delivered inside your repo and your process; overflow capacity for agencies and dev shops.
Proof: deemwar.com/work · deemwar.com/insights/second-model-reviewer
Make your product and team AI-native
Coding agents working in your repo on one real workflow first, with checks the agent did not write, cost tracking, and secrets masked before the model sees them.
Proof: deemwar.com/insights/sample-agent-secrets-audit · deemwar.com/explainer/code-review-without-prs (background)
Cloud, data and migration work
Any shape: infrastructure, pipelines, database moves, each verified before cutover.
Tell us your scope: book a 20-minute call.
Leave your work email or write to io@deemwar.com, and a person replies within a day.
Tell us your scope: book a 20-minute call.
1 week
“Everyone on the team uses Claude Code differently, and nobody knows what it can see.”
What we build. A team CLAUDE.md, hooks, skills and MCP servers set up in your repos, with secrets kept out of the model.
Proof. Before: we scanned 423 MB of our own coding-agent transcripts, and 72 files held secret-shaped values (an upper bound; it includes demo and test values). After: with Veilpipe set up, the model sees a name like $STRIPE_KEY, never the value.
Build + monthly run
“One back-office job eats hours every week, and nobody wants to do it.”
What we build. An AI agent that does that one job in your systems, with guardrails and a person to approve, and we run it for you every month.
Proof. Our own inbox, last 7 days: 148 of 221 inbound emails filed by the agent with no human; the rest handed to a person. (Our own back office; per-email timing; the model was capped for part of the week.)
Verified migration
“Our Postgres dashboards are slow, and we can't risk losing rows moving to ClickHouse.”
What we build. A snapshot plus live CDC into ClickHouse, with per-table checksums, a CI test that re-checks every change, and a cutover and rollback runbook.
Proof. In our CI demo, 15 tables seeded with 505,800 rows stayed equal by checksum after 5 minutes of live writes, and a dropped table was caught within 12 s.
Scoped per workload
“Our AWS, GCP or Azure bill keeps growing, and moving off feels too risky.”
What we build. Workloads and databases moved to Hetzner-class servers or your own hardware, across more than one region, with a planned cutover, tested backups and monitoring.
Proof. We run our own stack this way: Hetzner and Contabo servers joined over WireGuard. Database moves into ClickHouse are checked by checksum with chlift, our ClickHouse migration and mirroring tool. In our cutover test (Postgres 15 to 17, 505,800 rows seeded, then 5 minutes of live writes), the old database was read-only for 6.5 s at cutover, measured after the load ended, and the pre-cutover check caught a table left out and 14 sequences that would have broken the first insert.
“Our data copies drift silently, and we only find out when a report is wrong.”
What we build. Pipelines and migrations that verify themselves (row counts, checksums, alerts), plus backups you own and have restore-tested.
Proof. chlift: 0 of 64 buckets differ, equal after live changes (preview). pgbx: Postgres extension + CLI, tested on Postgres 13–18.
1 week
“We don't know where our AI agents can see or leak our API keys.”
What we build. A severity-ranked leak report across env, logs, transcripts, CI and MCP, plus secret brokering working in one of your repos.
Proof. We scanned our own agents' transcripts, and built the broker: Veilpipe, 232 patterns, a 65 s demo where every command really runs.
“Our coding agents run destructive commands and follow injected instructions.”
What we build. Small local yes/no checks next to your rules: prompt-injection, destructive-command and claim checks, running on CPU on your machines.
Proof. The local model caught 42/60 injections vs 20/60 for a keyword rule, and 16/20 destructive commands vs 1. The same person wrote parts of the test sets and the rules.
“We need a solid CLI or dev tool, and nobody on the team has time to build one.”
What we build. Cross-platform CLIs with installers for Linux, macOS and Windows, published and documented: open-source or for your internal use.
Proof. pgbx, chlift, jevx and AgentProxy, each with its own product page and installers.
2 weeks
“Our agent works in a demo but isn't production-ready: no evals, guardrails or cost tracking.”
What we build. One agent workflow in your repo and cloud: provider layer, tool calls, guardrails, evals with confidence intervals, logging and cost.
Proof. AgentProxy logs every agent model call with token counts and cost per agent, and stops an agent at a limit you set.
“We want an agent to run a workflow on its own, but not without limits it cannot break.”
What we build. The agent plus its guardrails, for ops, finance or support: hard limits checked before every action, a watchdog, a kill switch only a person can re-arm, and a log of every refusal with its reason.
Proof. Our own trading desk runs this way, 24/7: a watchdog every 60 s, a drawdown ladder, and 1,994 refusals logged with reasons in one archived log.
Workflow build
“Our agents do good work once, but we can't make the same job run the same way every time.”
What we build. One workflow from a recipe (production watch, alert triage, CI failure doctor, PR review): typed steps whose output must validate before the next one runs, human approval gates, cost per step.
Proof. We built our own workflow engine for this and run it in-house: one binary, self-hosted, works with Claude, Codex, Copilot, OpenCode or Devin.
“We want an AI agent doing real work in our product, not a chatbot demo.”
What we build. Agents with tool calls and guardrails, deployed in your cloud.
Proof. Embedded engineering partner for an AI company in the Netherlands (under NDA), plus our own open agent tooling.
“We need a working, paying product live, not a prototype.”
What we build. A full SaaS: auth, payments, background jobs, AI features and deploys, on a Go + Postgres + React stack we already run in production.
Proof. reqsume is our production SaaS (AI rewrites, ATS-tailored exports, credit-metered usage). For ACE Financials: full product ownership through to production, including GCP deployment.
“Our backend can't keep up: jobs get stuck, queues misbehave, the API is brittle.”
What we build. Go/Java backends, Postgres-backed job queues and APIs, plus your API as a CLI, agent skills and an MCP server.
Proof. The reqsume job queue runs on Postgres with no separate watchdog.
“We need a web app or an internal dashboard our team will actually use.”
What we build. React web apps with a Go/Postgres backend: customer-facing apps, admin panels and live dashboards.
Proof. reqsume is a live web app; AgentProxy shows every agent's calls and cost on one dashboard.
“Our deploys are manual and scary, and we're not sure the alerts would ever fire.”
What we build. One-command deploys, CloudFormation/AWS Marketplace packaging, CI pipelines and alerting tested end to end.
Proof. openjevx Server deploys to ECS Fargate from one CloudFormation stack; pgbx ships as an AWS Marketplace image.
“Bugs reach production because our tests check the wrong things.”
What we build. End-to-end and CI tests, browser checks, mutation testing and independent second-model review of agent-written code.
Proof. A browser check caught a UI bug of 19 stacked cards; chlift's CI end-to-end test re-verifies every change; we teach JUnit/Mockito and Stryker mutation testing.
“Our launch has a README, not a demo anyone will watch.”
What we build. Narrated demo videos of the real product running, with captions.
Proof. 21 videos on our developer-tools channel.
More projects, one card each, in the work gallery.
We move workloads and databases off AWS, GCP or Azure to Hetzner-class servers or on-prem hardware: more than one region, a cutover with no downtime, backups that have been restored at least once, and monitoring that has been seen to fire.
It is how we run our own company: Hetzner and Contabo servers joined over WireGuard. Each database copy is checked table by table before cutover. Tell us the workload in the form above and we scope it with you.
RAG over your documents with pgvector, AI rewrite and summary features, and evals that prove they work.
So far. So far in our own product (reqsume), not yet for a client.
A codebase assessment, a prioritised fix plan, then the fixes, with tests added as we go.
So far. We've done the assessment for Arus Innovation (codebase + notification-system architecture review).
Phone-call and SMS agents on a box with your own SIM.
So far. Early access: a 2-minute, 3-turn AI call on a live mobile network; boards designed, not yet made.
Chrome extensions, from content scripts to a published listing.
So far. A demo video so far; no store listing yet.
Tell us your scope: book a 20-minute call. io@deemwar.com
