AI strategy and implementation
for PE-backed companies
I partner with PE firms and their portfolio companies to surface profit-driving AI initiatives and ship them to production.
Problems I solve
- Initiative sprawl with no business caseScattered use cases with unclear ROI. You need the few that drive most of the value.
- Stuck at 80%Prototyping to 80% is easy. The last 20% is where projects quietly die.
- No systematic process to measure and improveEvery prompt, pipeline, or retrieval tweak is a coin flip. You cannot tell improvement from regression.
- An AI that explains but can't actIt answers questions but can't update the CRM, open the ticket, or trigger the workflow.
Services
- 01AI strategyIdentify productivity gaps and prioritise high-ROI use cases on an impact-effort matrix.
- 02AI implementationArchitect and deploy production AI systems that hold up under real-world scenarios.
- 03AI continuityPeriodic reviews after handover catch silent regressions and keep the system performing.
Background
- Bright Pixel Capital
- Lead AI Engineer. €600M B2B infrastructure software fund. Shipped production systems for core fund workflows and ran technical assessments on 100+ companies across devops, devtools, data, and AI infrastructure.
- Deloitte Digital
- Tech Consultant. Led digital transformation programmes for major commercial banks.
How I work
Every engagement starts with a 4–6 week pilot on one narrow problem. By the end, we've proven feasibility and quantified the business case for the full rollout.
- Forward-deployedI work as an embedded operator inside the team for the project duration.
- Evaluation-centricEvery system I ship has a measurement layer before it has a UI.
- Built for handoverYour in-house team can operate and extend the system independently when I leave.
Convictions
A set of principles learned the hard way from building production systems. If you don't relate, we're probably not a fit.
- 01Process over toolsMost production AI value lives in the system around the model, not in the model itself. If you ask me to swap the database or the LLM to fix a quality issue, I'll push back and start with the pipeline.
- 02Prompt and pipeline before fine-tuningMost failures aren't a model problem, they're a specification problem. If you want to fine-tune before we've exhausted prompt engineering and pipeline architecture, I'm not the right person.
- 03Systematic evaluation at the coreDevelopment and evaluation are one engagement, not two. If you want to scope evaluation as Phase 2 or cut it to save budget, I'll push back.
- 04Deterministic workflows over fully agentic systemsAgents are useful at specific points in a pipeline, not as a general solution. If you want a fully agentic workflow before we've mapped the actual process, I'll pass.
- 05Augment the decision, don't automate itThe highest-value AI systems reduce cognitive load around a decision, not the decision itself. If you want an AI that makes the high-stakes call autonomously, I'm not the right person.
Who I work with
Right fit
- PE-backed portfolio companies with operational drag in a specific domain.
- VC-backed scale-ups from Series B onwards, gated by operational throughput.
- Mid-market ops-heavy operators with meaningful revenue and margin.
Not the right fit
- Teams looking for a fully autonomous agent as a magic bullet.
- Companies without the revenue, margin, or organisational readiness.
- Companies building their own foundation models or novel architectures.
FAQ
How long does an engagement take?
Typically 3–6 months, with a 4–6 week pilot. By the end of the pilot, we've proven feasibility, quantified the business case, and defined the scope for the rest of the engagement.
How do we collaborate?
Weekly working sessions to review what shipped and what's next, shared visibility on the roadmap, and direct communication between sessions when something needs a fast decision.
What's the pricing model?
Two stages. The pilot is fixed-fee and produces the business case. The remaining engagement is priced against that case.
What are the deliverables?
A working system in production, the documentation, and the deployment instructions to run it. Handover is structurally built into the engagement.
How do you handle privacy and data?
Default to open-source, self-hostable infrastructure that can run in the client's cloud or on-prem. Client data is never used to train any third-party model.
What is the tech stack?
Python-first, containerised, built on open-source frameworks to avoid vendor lock-in across orchestration, retrieval, and data layers.
Which models do you use?
Frontier-model APIs from Anthropic, OpenAI, and Google by default. Open-weights models when self-hosting is in scope.
How is the system deployed?
Client's cloud or on-prem, depending on infrastructure and compliance constraints. The client owns the stack and can change providers as needed.
What if we need to extend the system?
Most engagements include a continuity layer for exactly this. Periodic reviews catch silent regressions, recalibrate as data drifts, and address new requirements.
If you want to move a metric or find one worth moving, get in touch. The first call is a working conversation about the problem and the fit.
Book a callAbout
Most AI consultants are either engineers who can't translate the work into business value, or business people who can't ship code. I do both.
I'm Rui Sá, a senior operator with a decade across applied AI, software engineering, and tech consulting. Find me on LinkedIn.