Case studies

Real systems. In production.

Three systems we've designed, built, and operated. Clients stay unnamed – the metrics and technical details are exact.

IT Distribution In production – daily use

#01

Enterprise financial orchestration: 2–3 hours of CFO time freed daily

The challenge

The finance function was losing 2–3 hours of CFO time daily to manual vendor reconciliation across multiple major vendors. Correspondence, invoices, statements, and supporting documents all arrived in different formats and were processed by hand before anything reached the ERP. Reporting cycles were manual end to end.

Our approach

We designed and delivered an AI orchestration system that runs in production every working day. Agents read and draft vendor correspondence, reconcile invoices against orders and statements, classify inbound documents, and sync clean records to the ERP – with every exception escalated to a human. LangGraph coordinates the agents; FastAPI services on PostgreSQL hold state; n8n moves data between systems; everything ships in Docker. Releases are gated on evals: reconciliation matches are scored against a golden set with known-correct answers, document classification is checked against a labelled corpus, and the suite re-runs on every prompt or model change – a regression stops the deploy.

Results

2–3 hrs

CFO time freed

every working day

Eliminated

Manual reporting cycles

replaced by automated sync

Multi-vendor

Vendor coverage

across multiple major vendors

LangGraphn8nFastAPIPostgreSQLDocker

Voice AI / Sales Live system – outbound SDR in production

#02

PulseSales: a deterministic real-time voice agent for outbound sales

The challenge

Outbound SDR work – qualification calls, appointment confirmations, follow-ups – is repetitive, time-boxed to business hours, and hard to staff consistently across languages. Free-form LLM calls are unpredictable; sales conversations need controlled, repeatable flows.

Our approach

We built a real-time voice agent on the OpenAI Realtime API over WebRTC. A 13-state workflow engine controls every conversation – greeting, qualification, objection handling, booking, wrap-up – so behaviour stays predictable while the voice stays natural. A FastAPI/PostgreSQL backend manages state and outcomes; a Next.js 15 dashboard shows live calls, transcripts, and results. Every release passes a state-path eval suite: scripted transcripts must land in the correct terminal state, and no call reaches qualification without passing the disclosure state first.

Results

13-state

Conversation workflow

deterministic flow control

Multilingual

Outreach

outbound SDR + confirmations

Real-time

Voice transport

over WebRTC

OpenAI Realtime APIWebRTCFastAPIPostgreSQLNext.js 15

Operations Automation In production – continuously extended

#03

Multi-agent operations platform: daily ops cut from ~2 hours to under 15 minutes

The challenge

Routine operations – triaging inboxes, monitoring leads, compiling KPI reports, running scheduled checks – consumed around two hours of focused time every day. Running always-on agents through paid APIs would have made per-task costs unpredictable as the agent count grew.

Our approach

We built an orchestration platform on Linux and Docker running 30+ specialised agents on local models: inbox triage, lead monitoring, KPI reporting, and ops automation, with scheduling, retries, and human escalation paths. Local inference means zero marginal LLM cost – agents can run as often as the work demands. Changes ship behind evals: triage routing is scored against a labelled email set, generated KPI figures are recomputed from the source database to catch hallucinated numbers, and replayed inputs must produce no duplicate side effects.

Results

30+

Agents in production

running daily

<15 min

Daily operations time

down from ~2 hours

Zero

LLM cost

local models – no per-token spend

Multi-agent orchestrationDockerLinuxLocal LLMs

05Start here

Tell us what you're trying to automate

Not sure which service fits? Describe the bottleneck, the pilot that stalled – or the automation your last developer left behind – and we'll map the right system and scope it honestly.

Start a conversation →Browse services & pricing