← Back to home
Selected work.
Three projects. Each one replaced a rented stack or a stack of hands. Below: the audit, the build, the numbers before and after — and what it actually cost.
Case · 01 — SaaS · Logistics tech
The agent that learned the product.
Tier-1 support went from two FTEs drowning in tickets to one agent resolving 73% of them, autonomously, in under a minute.
Live
73%
tickets auto-resolved
47s
median time to resolve
2 FTE
redeployed to retention
The situation
- Two support agents spent 6 hours/day on Tier-1 (order status, ETA, password resets, BOL retrievals).
- Off-the-shelf chatbot answered 9% — it didn't know their products, accounts, or policies.
- CSAT was sliding: median first-reply 4h20.
Cost of the status quo · €84k/year in salary spent on copy-paste answers.
Engagement
- Scope · AI support agent · CRM-aware
- Client · B2B SaaS · 200+ customers
- Team · 18 people · 2 in support
- Built in · 9 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1
Audit
Shadowed support for 3 days. Tagged 600 tickets by intent. Mapped which 11 intents covered 78% of volume.
Week 2
Blueprint
Decision tree per intent. Defined what the agent is allowed to write back to the CRM, what it must escalate, and the refund authority ceiling.
Week 3–7
Build
Agent wired to their Postgres + CRM via read/write tools. Policy guardrails. Eval suite with 240 golden tickets.
Week 8
Shadow run
Agent answered in parallel with humans. We compared every response. Tuned 14 edge cases.
Week 9
Handover
Cut over. Live dashboard for the team. Runbook + retraining flow they own.
What we shipped
- Intent classifier trained on 600 real tickets.
- Tool layer: orders, accounts, BOLs, refunds (capped).
- Escalation router with reason codes.
- Eval suite they run before every prompt change.
Postgres (read/write)Their CRM APIGPT-5 + guardrailsEval harnessSlack alerts
Before · After
Tier-1 resolution rate
9%
→
73%
Median first response
4h 20m
→
47s
FTE hours/day on Tier-1
12h
→
1.5h (supervision)
CSAT (rolling 30d)
3.8
→
4.6
ROI
€71k/year
Year 1 net — net of build cost.
"It answers things our junior support team used to ask me about. And it never sleeps on a Sunday."— Head of Operations, B2B SaaS · 200+ customers
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Case · 02 — Wealth · Independent AM
One panel retired the stack.
Three SaaS subs and two days of monthly copy-paste collapsed into one internal panel. Reports now ship in an hour.
Live
3 subs
cancelled
16h → 1h
monthly report cycle
€38k
annual SaaS removed
The situation
- HubSpot + Airtable + a dashboard tool, each holding a slice of the truth.
- Two full days every month copy-pasting between them to produce the client PDF.
- Nobody trusted the numbers — three sources, no reconciliation.
Cost of the status quo · €38k/yr in SaaS + 24 person-days/yr in manual reporting.
Engagement
- Scope · Internal ops + reporting tool
- Client · Asset manager · Zürich
- Team · 18 people · 3 on reporting
- Built in · 6 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1
Audit
Mapped every field across the three tools. Identified what was actually used vs. legacy noise. Killed 60% of fields outright.
Week 2
Blueprint
One canonical schema. Custodian feed as source of truth. Defined the exact monthly report template with the partners.
Week 3–5
Build
Internal app with auth, role-based views, one-click PDF export, audit log.
Week 6
Handover
Migration of 4 years of historical data. Subs cancelled on a calendar. Training for the three reporting users.
What we shipped
- Single data layer pulling from custodian nightly.
- Templated client report — generated, not assembled.
- Partner dashboard with AUM, flows, allocation drift.
- Audit log on every export.
Custodian APIPostgresTanStack + SupabaseServer-side PDFRole-based auth
Before · After
Monthly report cycle
16 hours
→
1 hour
Active SaaS subs
3
→
0
Source of truth
3 places
→
1
Annual SaaS spend
€38k
→
€0
ROI
€58k/year
Year 1 net — net of build cost.
"The partners trust the numbers again. That's the part you can't put on an invoice."— Managing Partner, Asset manager · Zürich
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Case · 03 — Logistics · 3PL
PDFs in. Clean records out.
Hundreds of POs, BOLs and supplier confirmations a day, transcribed by hand. Now they flow into the ERP with 12 minutes of human supervision.
Live
14h → 12min
daily transcription
99.4%
field accuracy
€41k
errors eliminated / yr
The situation
- Hundreds of POs, BOLs and supplier confirmations arrived as PDFs and email attachments daily.
- Two ops people spent their first 3 hours each morning transcribing into the ERP.
- Five-figure mistakes happened twice a quarter from a misread digit.
Cost of the status quo · €62k/yr in salary + ~€41k/yr in error remediation.
Engagement
- Scope · Document automation agent
- Client · Logistics ops · Bucharest
- Team · 42 people · 2 on data entry
- Built in · 7 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1
Audit
Pulled 400 historical documents. Categorised by supplier and layout. Found 12 templates covered 91% of volume.
Week 2
Blueprint
Extraction schema per document type. Defined the 'flag for human' rules and the ERP write contract.
Week 3–6
Build
Vision + structured extraction. Validation against ERP master data. Human-in-the-loop queue for the 6% that needs a second pair of eyes.
Week 7
Handover
Two-week shadow period. Weekly accuracy report. Training the ops team to own the review queue.
What we shipped
- Inbox watcher (email + shared drive).
- Per-template extractor with confidence scores.
- Master-data validator (suppliers, SKUs, prices).
- Review queue for low-confidence rows only.
Vision LLMERP REST APIPostgres queueSchema validatorDaily accuracy report
Before · After
Daily transcription time
14 hours
→
12 minutes
Field accuracy
~97%
→
99.4%
Errors > €1k / quarter
2
→
0 (12 months)
Documents processed/day
~180
→
~340
ROI
€89k/year
Year 1 net — net of build cost.
"We doubled volume without hiring. The ops team finally has time for exceptions, not typing."— COO, Logistics ops · Bucharest
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Case · 04 — Professional services · Design
A portal clients actually log into.
Replaced a chaos of email threads, Dropbox links and WhatsApp groups with a white-label client portal. Status, files, approvals — one URL.
Live
92%
of clients log in weekly
−68%
status-update emails
11 days
faster average approval
The situation
- Every project lived across email, Dropbox, WhatsApp and a shared Notion. Nothing was the source of truth.
- Project leads spent ~5h/week answering 'where are we?' from clients.
- Approvals on renders and drawings dragged 2–3 weeks because no one knew whose turn it was.
Cost of the status quo · ~€72k/yr in partner time spent chasing approvals and re-sending files.
Engagement
- Scope · White-label client portal
- Client · Architecture studio · Milan
- Team · 24 people · 6 project leads
- Built in · 8 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1
Audit
Interviewed 6 project leads and 4 clients. Mapped the actual milestone flow vs. what the playbook said.
Week 2
Blueprint
Defined the canonical project state machine: brief → concept → DD → CD → site. One approval primitive across all of them.
Week 3–7
Build
Multi-tenant portal, role-based access (client / consultant / internal), file versioning, threaded approvals, email + portal notifications.
Week 8
Handover
Migrated 14 active projects. Onboarded 40 client contacts. Internal champion trained on configuration.
What we shipped
- Per-project workspace with status, files, approvals, comments.
- Approval primitive with deadlines and reminders.
- Client-side view that hides internal noise.
- Audit trail on every file and decision.
TanStack StartPostgres + RLSS3-compatible storageResend (email)Audit log
Before · After
Avg. approval cycle
18 days
→
7 days
Status emails / project / week
~22
→
~7
Source of truth for files
Dropbox + email
→
Portal
Client weekly active rate
n/a
→
92%
ROI
€84k/year
Year 1 net — net of build cost.
"Our clients stopped asking. That alone gave each project lead a day back per week."— Founding Partner, Architecture studio · Milan
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Case · 05 — Consulting · 40-person firm
Quote to cash, in one pipeline.
Killed HubSpot + PandaDoc + a spreadsheet by building one pipeline: lead in, signed contract and invoice out, no copy-paste in between.
Live
−€31k
in SaaS cancelled
9 days → 2
lead to signed contract
+18%
close rate on warm leads
The situation
- Leads sat in HubSpot, contracts in PandaDoc, invoicing in a spreadsheet, payments in Stripe. Nothing reconciled.
- Sales ops re-keyed every deal 3 times. Errors leaked into invoices.
- Average lead aged 9 days before a contract went out — most lost momentum.
Cost of the status quo · €31k/yr in SaaS + 30+ person-days/yr re-keying deal data.
Engagement
- Scope · Sales pipeline + contracting
- Client · B2B services · Berlin
- Team · 40 people · 5 in sales ops
- Built in · 10 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1–2
Audit
Sat with sales for a week. Mapped every status, every handoff, every field actually used. Killed 40% of the CRM schema.
Week 3
Blueprint
One deal object from lead to paid. Defined the contract template engine and the Stripe handoff.
Week 4–8
Build
Pipeline UI, contract generator with e-sign, Stripe invoice + payment links, revenue dashboard for partners.
Week 9
Shadow run
Ran new pipeline in parallel with HubSpot for 10 days. Reconciled every deal. Fixed 6 edge cases.
Week 10
Handover
Cancelled HubSpot + PandaDoc. Trained sales + ops. Quarterly review cadence established.
What we shipped
- Single deal object with full status timeline.
- Contract template engine — auto-filled from the deal.
- E-sign + Stripe payment link in one email.
- Partner revenue dashboard with forecast.
PostgresTanStack StartStripe (invoices + checkout)DocuSign-compatible e-signResend
Before · After
Lead → signed contract
9 days
→
2 days
Tools in the stack
4
→
1
Manual re-keying / week
~6 hours
→
0
Close rate (warm)
34%
→
52%
ROI
€96k/year
Year 1 net — net of build cost.
"We stopped paying four vendors to do half a job each. The pipeline is finally one thing."— Managing Director, B2B services · Berlin
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Case · 06 — HVAC · Field operations
The dispatcher that plans the day.
Replaced a whiteboard and a phone tree with a dispatch tool that assigns jobs, routes vans and texts customers — before the dispatcher's coffee is cold.
Live
+22%
jobs per technician / day
−43%
missed appointment rate
9.1/10
customer NPS post-visit
The situation
- Two dispatchers planned the day on a whiteboard, by phone, every morning from 6:30.
- Customers got a 4-hour window and a phone call 'on the way'. Often they weren't home.
- Route inefficiency burned ~80km/day per van in detours.
Cost of the status quo · ~€118k/yr in lost technician hours, fuel and missed-visit re-scheduling.
Engagement
- Scope · Dispatch + scheduling tool
- Client · Field services · Vienna
- Team · 60 people · 28 technicians
- Built in · 11 weeks
Fixed scope · fixed price · live deliverables every Friday.
How we audited & built it
Week 1–2
Audit
Rode along on 12 jobs. Tracked actual time on site vs. quoted. Found the planning, not the work, was the bottleneck.
Week 3
Blueprint
Defined the assignment algorithm (skills, location, SLA), the customer comms cadence, and the dispatcher override model.
Week 4–9
Build
Dispatcher board, technician mobile app with offline mode, customer SMS flow with live ETA, route optimisation per van.
Week 10
Shadow run
Two weeks in parallel. Dispatcher had veto on every auto-assignment. We tuned the algorithm based on every override.
Week 11
Handover
Cut over. Whiteboard retired. Weekly metrics review with ops manager.
What we shipped
- Auto-assignment with dispatcher override.
- Technician PWA — works in basements with no signal.
- Customer SMS with live ETA and 'I'm 15 min away'.
- Daily KPI board: jobs/tech, miss rate, route efficiency.
Postgres + PostGISTanStack StartTwilio (SMS)MapboxOffline-capable mobile PWA
Before · After
Jobs per technician / day
4.1
→
5.0
Missed appointments
11%
→
6%
Daily planning time
2h × 2 people
→
20 min × 1
Fuel cost / month
€8.4k
→
€6.1k
ROI
€142k/year
Year 1 net — net of build cost.
"The dispatchers now manage exceptions, not the whole day. Technicians do one more job each. It compounds."— Operations Director, Field services · Vienna
Have a similar problem? We can scope a build like this in a week.
Request a similar solution →Yours is the fourth case.
We're booking projects for Q2 2026. One short call, one written scope, fixed price. If we're not the right fit, we'll tell you in the first 20 minutes.
Start the conversation →