We work with operators and engineering leaders to take ambiguous business problems and turn them into production-grade systems — orchestrated, observable, reversible. Not pilots. Not demos. Systems your team can run on a Sunday at 03:00 and on quarter-close at 17:30.
Each capability below is something we design, ship and operate. They compose into one system with one event log, so every agent step, tool call and approval can be traced and replayed.
Planning, retrieval, tool use and evaluation in one traced run. Humans approve where policy says they must.
Scoped MCP servers over internal systems, with audit on every call.
Intake and triage calls that capture structured data and hand off cleanly.
Hybrid search and reranking, tuned against a labelled recall set.
Offline suites plus online sampling. Regressions block the deploy.
Ingest, reasoning, action, approval, audit and observability run as one system with one contract.
OpenTelemetry spans on every model call, tool call and approval.
Most teams stitch together SaaS tools with people, spreadsheets, and Slack threads. It works at small scale and shatters at the seams. We rebuild the seams as systems.
Eight SaaS systems, three spreadsheets, and a Slack channel where the handoff actually happens. Status is folklore. Audit is grep. Coverage drops at 5pm on Friday.
A single event-sourced core mediates every step. Every action is observable, reversible, and replayable. The same UI runs the live system and the post-mortem.
Most projects are a remix of a small number of well-trodden patterns. Filter the graph to see which capabilities, tools and industries each one touches.
The maturity ladder is a real progression — most teams sit between L2 and L3. We rebuild operations as orchestrated workflows that audit themselves.
Most AI projects fail at the layer below the model. We build all seven — and refuse to call anything done until the evaluation layer says so.
Architecture is a deliverable. Source systems feed one orchestration core; the core is the only place anything is decided; every product surface reads from it.
We work across financial services, supply chain, healthcare ops, and customer-facing software. The patterns rhyme — the implementation always differs.
The work isn't done when the demo works. We ship with the operational scaffolding production actually requires — and we won't bill the milestone without it.
Target ranges for the kind of work we take on, measured in the metrics operations teams already report. Illustrative · target operating model.
We're senior generalists by design — but the work splits cleanly into four engagement types. Most projects combine two or three of these.
We sit in your operation, map the system as it really runs, and return with an event model, a target architecture, and a phased build plan you could hand to any team.
Founder-led senior engineering ships the system end-to-end. Weekly demo, paired with your engineering, with production readiness as the bar.
We continue to operate the system with you — on-call rotation, eval cadence, model upgrades, integration drift. Until you don't need us.
For teams that want to build it themselves. A senior engineer embedded part-time as architect-of-record, with reviews, paired sessions, and on-demand pairing.
The shape of the engagement is always the same. The work inside it never is. Most projects go from intake to production in 8–16 weeks.
The tools we reach for first. The project picks the tools, not the team. Hover a tool to find it in the graph. No vendor logos.
A founder-led senior studio — hands-on engineers who got tired of pilots. We write more than we tweet, and we'd rather lose a deal than ship a bad system.
Soft Actions is founder-led: the person who scopes your system is the one who builds it. The work spans LLM agents and workflow automation, real-time speech systems, quant and trading infrastructure, and data-heavy SaaS platforms. The studio runs on the same principles the systems do — small, senior, paired, audited.
No open roles right now. We sometimes bring in senior engineers for specific projects — if that is you, send a short note through the brief form below.
"We design the dashboard before the model. We design the runbook before the dashboard. We don't take a job we couldn't sit in the on-call rotation for." — SOFT ACTIONS
"We design the dashboard before the model. We design the runbook before the dashboard. We don't take a job we couldn't sit in the on-call rotation for."
Most engagements start with a brief. Tell us what you're trying to operate, where it hurts, and what you've tried. We respond within two business days.
Discovery sprints are fixed-price. Build engagements are outcomes-based with a price range agreed up front and re-negotiated at the prototype gate. We don't do open-ended time-and-materials engagements above 8 weeks.
If the problem fits in one 8-week build, that's probably the right shape. Below that, we recommend a 2-week discovery sprint and a referral to a smaller team.
Yours. We add Soft Actions engineers to your repository, your tooling, your CI. We don't maintain a parallel infrastructure that becomes a handover tax.
You own all code, IP, and event data on day one. Soft Actions retains the right to reference engagement shape and outcomes (anonymized) for marketing. NDAs signed as standard.
Yes — financial services, healthcare ops, and adjacent. We build to SOC 2 / ISO 27001-style controls — audit logs, access control, PII handling — and adapt to each client's own requirements.
Yes, with conditions. We need 2 weeks of read-only review before committing — most rescue work is misdiagnosed and we won't take it on without a clear failure mode in mind.
Remote-first, working with clients wherever they are. Most engagements run fully remote, with on-site days at kickoff and stabilization when a project needs them.
The ones that come up on the first call, answered up front.
If your operation is held together by people, spreadsheets and Slack — and the seams are showing — we'd like to hear about it. The first call is 90 minutes. The second is a written discovery memo.