Case studies · Field traces
Systems designed, built and operated.
Systems designed, built or operated under my responsibility, grouped by family: acquisition, data, voice agents, orchestration. Client names remain confidential by default because these engagements often touch business flows, proprietary know-how, data and internal trade-offs.
Each entry documents what can be useful without exposing the organisation: architecture, primitives, governance, limits, human takeover and observed effects.
Two co-built systems were recognised at the HighLevel AI Agents Contest 2025: a conversational agent awarded for real-world application, and a voice agent finalist for technical excellence.
How to read this
Proof does not come from logos.
In this kind of work, showing a client name is not always the strongest signal. AI systems often touch business flows, data, know-how and internal trade-offs.
This page documents proof differently: through the problem addressed, the chosen architecture, guardrails, observed effects and the level of human control preserved.
01
Anonymised case
Real system, described without names, sensitive data or context specific enough to expose the organisation.
02
Abstracted pattern
Real problem encountered or observed, reformulated to transmit the pattern without revealing the company, exact sector or IP.
03
External proof
Public or authorised signal: contest, ranking, published artifact, shareable feedback or metric already disclosed in a compatible context.
Documented systems
Jump directly to the system you want to inspect.
Seven systems are documented below. Each entry keeps the same reading frame: context, design, governance, outcome and primitives.
01 · Multi-Agent · Chat
24/7 Multi-Agent Chat Booking
WinnerHighLevel AI Agents Contest 2025Conversation AI · real-world application
Autonomous multi-agent chat system that qualifies visitors, answers questions, handles objections and drives appointment booking 24/7. Built for business services with human-grade precision requirements.
Role · Architect · Build · Run
System map
External proof
From signal to operated output.
01
Visitor
02
Qualification
03
Objections
04
Booking
05
Escalation
Humans take over ambiguous cases and validate business rules.
Agent mandate · Qualification contract · Objection ledger · Booking handoff
- Context
- Inbound acquisition uncovered at night and weekends; recurring objections poorly handled by static chatbots; sales calendar saturated with unqualified conversations.
- Design
- Split into specialised agents (greeting, qualification, objection handling, booking) with explicit interaction contracts. Short session memory, persistent business memory for product rules, typed human escalation on doubt signal.
- Governance
- Each booked appointment is traced with the conversation that drove it. Product rules live in a governed store, not in the prompt. Editorial changes go through review before deployment.
- Outcome
- Inbound acquisition stabilised outside business hours. Sales calendar filled without manual work. Short editorial review cycle.
Primitives
- Agent mandate
- Qualification contract
- Objection ledger
- Booking handoff
- Escalation signal
Observed in production
- 40–120Qualified appointments / month
- 50–70 %Chat workload reduction
- 24/7Continuous qualification
02 · Multi-Agent · Voice
24/7 Multi-Agent Voice AI Ops
FinalistHighLevel AI Agents Contest 2025Voice AI · technical excellence
Natural-language voice agents that answer inbound calls, route requests, handle support, book appointments and escalate to a human when context requires it.
Role · Architect · Build · Run
System map
External proof
From signal to operated output.
01
Incoming call
02
Intent
03
Action
04
Transcript
05
Support or booking
Humans keep sensitive commitments and complex takeovers.
Voice envelope · Intent classifier · Action capability · Escalation route
- Context
- Saturated phone reception, missed out-of-hours calls, heterogeneous answer quality across operators, response cost linear to volume.
- Design
- Real-time voice loop split into agents (greeting, intent, action, escalation). Sub-second target latency. Explicit boundary between autonomous actions (lookup, booking) and sensitive actions (contractual commitment, payment).
- Governance
- Each call leaves an indexed transcript. Sensitive actions are never executed without human validation or explicit contract. Latency and quality measured by cohort.
- Outcome
- No more lost calls outside hours. First-line support largely autonomised. Human escalation concentrated on cases worth the effort.
Primitives
- Voice envelope
- Intent classifier
- Action capability
- Escalation route
- Replay key
Observed in production
- 70–90 %Response time acceleration
- 30–50 %Support workload reduction
- 24/7Qualification and booking
03 · Lifecycle · Human + AI
Hybrid Human + AI Appointment Lifecycle
Full lifecycle: booking, reminder, meeting, bilingual recap, follow-up, inbound triage. AI orchestrates, humans validate committing actions.
Role · Architect · Build · Run
System map
Anonymised case
From signal to operated output.
01
Booking
02
Reminder
03
Meeting
04
Recap
05
Follow-up
Humans validate committing messages and relationship trade-offs.
Lifecycle event bus · Bilingual recap · Follow-up scheduler · Inbound triage
- Context
- Firms and executives overloaded by relationship admin: manual recaps, forgotten follow-ups, untriaged inbound, scattered notes. Premium experience promise hard to keep.
- Design
- Event-driven pipeline: each lifecycle milestone triggers an agent. Automatic bilingual recap (FR/EN), scheduled follow-ups, AI triage of inbound forms with human-validated email drafts.
- Governance
- No committing email is sent without human review. Every step is auditable. Sensitive wording is templated and versioned.
- Outcome
- Zero forgotten follow-up. Predictable, premium client experience. Drastically reduced admin load on the team side.
Primitives
- Lifecycle event bus
- Bilingual recap
- Follow-up scheduler
- Inbound triage
- Human draft validation
Observed in production
- 0Forgotten follow-up
- FR/ENInstant bilingual recap
04 · Media · Source-to-Artifact
Automated AI Video Production Engine
URL → video → social publications pipeline. A web source becomes a publish-ready media format in minutes, without manual intervention.
Role · Architect · Build
System map
Anonymised case
From signal to operated output.
01
URL
02
Extraction
03
Script
04
Voice
05
Formats
Humans remain final editors on sources, rights and sensitive segments.
Source asset · Script artifact · Voice synthesis · Format profile
- Context
- Manual video content production, slow, costly, heterogeneous across brands and formats. High demand for multi-platform output, low execution capacity.
- Design
- Agent chain: source extraction, scripting, voice, editing, per-platform formats, scheduling. Each step produces a validatable artifact.
- Governance
- Human remains final editor on committing segments. Rights, sources and attributions are tracked at each step.
- Outcome
- Output capacity multiplied. Editorial consistency preserved. Time-to-publish reduced by an order of magnitude.
Primitives
- Source asset
- Script artifact
- Voice synthesis
- Format profile
- Publish slot
Observed in production
- 95 %Production acceleration
- Multi-plateformeOutput by default
05 · Platform · Multi-tenant
Multi-Tenant AI Content Platform
Multi-tenant SaaS platform where each client organisation gets its own space, brand voice and governance rules on a shared API-first architecture.
Role · Architect · Build
System map
Operated architecture
From signal to operated output.
01
Shared core
02
Tenant
03
ACL
04
Config
05
Observability
Humans govern rights, exports and capability changes.
Tenant scope · Capability ACL · Versioned config · Tenant observability
- Context
- Need to serve multiple clients with distinct editorial, legal and operational requirements, without duplicating the stack.
- Design
- Strict tenant isolation, granular capabilities, versioned configurations, common observability. Agency tenants vs client sub-accounts distinguished structurally.
- Governance
- No data crossover between tenants. Each capability change is tracked. Exports respect tenant scope.
- Outcome
- Fast onboarding of new tenants. Single maintenance for the common stack. Customisation without dedicated code branches.
Primitives
- Tenant scope
- Capability ACL
- Versioned config
- Tenant observability
06 · CRM · Lifecycle
Google Reviews & Customer Re-activation
Google Reviews solicitation and client re-activation system articulated on the existing CRM base. Per-segment personalisation, follow-up governance, consent respected.
Role · Architect · Build
System map
Anonymised case
From signal to operated output.
01
CRM
02
Segment
03
Consent
04
Sequence
05
Reactivation
Humans define the legal frame, cadence and sensitive messages.
Segment scope · Solicitation contract · Consent ledger · Repeat signal
- Context
- Rich but underused CRM base. Insufficient Google reviews for local SEO. Manual, inconsistent reactivation.
- Design
- Behavioural segmentation, typed solicitation scenarios, respectful cadence, tracked opt-out, measurement of effect on rating and repeat.
- Governance
- No solicitation without clear legal basis. Capped cadence. Instant opt-out, traced and respected across all channels.
- Outcome
- Strengthened local reputation. Predictable reactivation of dormant clients. Measurable effect on rating and repeat business.
Primitives
- Segment scope
- Solicitation contract
- Consent ledger
- Repeat signal
07 · Multi-Agent · Unified
Unified Multi-Agent Voice + Chat Ops
Unified voice + chat front where agents share common memory, common rules and coherent escalations, whatever the inbound channel.
Role · Architect · Build · Run
System map
Operated architecture
From signal to operated output.
01
Voice
02
Chat
03
Shared session
04
Rules
05
Handoff
Humans receive escalations with shared memory and channel context.
Client session · Cross-channel identity · Shared rule store · Channel adapter
- Context
- Voice and chat channels operated in silos, fragmented client memory, inconsistent experiences depending on channel, escalations that restart from zero on each reply.
- Design
- Unified client session layer, stable cross-channel identifiers, shared business rules, common voice+chat observability. Channel-specific agents attach to this layer without duplicating it.
- Governance
- Single decision grid, single operational memory, single escalation policy. Channel divergences are configurations, not forks.
- Outcome
- Client is recognised from one channel to the next. Human operators arrive in the escalation with full history. Business rules change in a single place.
Primitives
- Client session
- Cross-channel identity
- Shared rule store
- Channel adapter
- Unified trace
08 · Données · Observabilité
Unlocking data trapped in a closed tool
An ageing line-of-business ERP with read-only access: the customer data existed but stayed unusable. A capture layer made over 10,000 events a month usable across three sites, at a fraction of the cost of a generic orchestrator.
Role · Architecte · Intégration
System map
Documented case
From signal to operated output.
01
Signal
02
Agents
03
Control
04
Trace
05
Output
Humans keep decisions that commit the real world.
Event capture · Event normalisation · Customer segmentation · Read-only boundary
- Context
- A network of family leisure parks (three sites) ran on a niche-specific ERP: solid for ticketing and point of sale, but ageing, read-only API, with no serious email or SMS capability. Leadership had chosen to keep that tool.
- Design
- Capture across every available output (webhooks and read API), normalisation of events into an orchestration layer, then a clean, reusable, segmented customer base. The system captured beyond the validated scope, so further analysis could come later without rebuilding anything.
- Governance
- A strict directive from leadership: keep the ERP, do not replace it. The processing scope was defined and approved upfront. The data stayed the company's own, ready to plug into the mailing or reporting tool of its choice.
- Outcome
- Data that had been locked away became usable for segmentation and follow-up. Processing cost in the range of one hundred fifty to two hundred dollars a month, where a generic orchestrator would have run around four thousand euros a month.
Primitives
- Event capture
- Event normalisation
- Customer segmentation
- Read-only boundary
Observed in production
- 10 000+events / month
- 3sites
- ~-95 %processing cost
09 · Architecture · Réseau
A duplicable system, delivered in a month
The operator of a network of around forty gyms wanted to test a new model: independent coaches selling their services inside the network's gyms. Designed as a duplicable building block, delivered, tested and handed over in a month.
Role · Architecte · Build
System map
Documented case
From signal to operated output.
01
Signal
02
Agents
03
Control
04
Trace
05
Output
Humans keep decisions that commit the real world.
Duplicable snapshot · External direct debit · E-signed quotes · Commission ledger
- Context
- Making the gym, the independent coach and the client work together cleanly requires clear rules: who collects payment, who earns which commission, how a quote is signed, how a failed payment is recovered. Tested on two to three pilot gyms before rollout.
- Design
- A hyper-customised template designed as a snapshot: one account per site, duplicable on demand. Direct-debit payments handled by an external service (subscriptions, dunning, failures), personalised e-signed quotes, and a back office aggregating each party's commissions.
- Governance
- A deliberately bounded engagement: prototype, deliver, document, hand over. Documentation and training videos were left so the network could run and evolve the system in-house.
- Outcome
- A new revenue model made operable and a system ready to duplicate across the network. The capability was internalised, as the client wanted.
Primitives
- Duplicable snapshot
- External direct debit
- E-signed quotes
- Commission ledger
Observed in production
- 1 moisdesign → delivery
- ~40gyms targeted for duplication
10 · Agent vocal · Privacy
Surveying a profession, keeping nothing
An anonymised survey across a healthcare profession organised as a national network. Collection handled by a voice agent designed privacy-first: no transcript kept, only structured, anonymised summaries.
Role · Architecte · Build
System map
Documented case
From signal to operated output.
01
Signal
02
Agents
03
Control
04
Trace
05
Output
Humans keep decisions that commit the real world.
Interview agent · Structured extraction · Privacy by design · Call parallelisation
- Context
- Participants, volunteers interested in the final report, had signed a consent form. Their answers had to be collected at scale, by phone, without spending months on it.
- Design
- A voice agent ran the interview, asked the questions and adapted to each conversation. Useful information was extracted, structured, then summarised to feed the final study report. Several calls ran in parallel, with number rotation to scale volume.
- Governance
- The principle was set upfront and explained at the start of each call: no transcript kept, nothing superfluous stored. The database held only structured summaries of the points sought. Real analysis of the conversation, not keyword extraction.
- Outcome
- People not at ease with such technology agreed to answer an AI, because the framing was clear, respectful and useful. Early proof that agentic collection can be both effective and privacy-respecting.
Primitives
- Interview agent
- Structured extraction
- Privacy by design
- Call parallelisation
Observed in production
- 0transcripts kept
- Parallèlesimultaneous calls
11 · Acquisition
From a few quotes a month to a full pipeline
A small quote-based services business, barely visible online, was missing inbound demand for lack of follow-up. A measured acquisition chain: get found, capture every request, sign, follow up. In under ninety days, the pipeline filled.
Role · Architecte · Build · Run
System map
Documented case
From signal to operated output.
01
Signal
02
Agents
03
Control
04
Trace
05
Output
Humans keep decisions that commit the real world.
Lead capture · E-signature · Automated follow-up · Ad tracking
- Context
- Few inbound requests, no credible site or Google presence, and the rare requests went unanswered and un-followed-up. The problem was not the quality of the work, it was everything upstream.
- Design
- A conversion-focused new site, a well-built Google business profile, and the whole chain orchestrated: request capture, e-signature, confirmation, follow-up and reactivation emails. A small, tracked ad budget to prime the flow.
- Governance
- A bounded, measured ad budget, paid back by the return. Every request tracked and followed up automatically, with the human keeping control of the quote itself.
- Outcome
- In under ninety days, the business went from a few requests a month to several a day. Its pipeline is now full.
Primitives
- Lead capture
- E-signature
- Automated follow-up
- Ad tracking
Observed in production
- < 90 jto fill the pipeline
- mois → jourrequest cadence
Evidence layer
Signals are published by architecture, not by client logo.
Some traces can be shown publicly, others remain private because they touch business flows, data or client know-how.
01 · contest
Go High Level AI Agent Contest 2025
Joint participation with Remy around three submitted agents: two multichannel text agents and one multichannel, multi-agent voice agent.
External signal showing the practice was exposed to a jury, real constraints and international comparison.
02 · field
Intensive MCP / Go High Level testing
Active participation in a beta phase around controlling Go High Level accounts through agentic interfaces, APIs, permissions, confirmations and governance.
Strengthens credibility on the shift from an agent that talks to an agent that acts in a business environment.
03 · systems
Internal agentic OS and documented systems
Charlie OS, Reveal System, editorial workflows, multi-agent orchestrations and parallel work systems act as testbeds before publication.
Shows the doctrine comes from building, not from theoretical monitoring.
From proof to case
A system analysis only matters when it becomes operable.
For a concrete AI subject, the guided Reveal journey qualifies context before any meeting. Classic contact stays available for direct exchanges.