P8 · Program
Technology Consulting
In-house engineers and applied-AI practitioners who have shipped at Fortune 500 scale audit your architecture, AI substance, unit cost and security posture — then tell you the truth about scale, margin and defensibility.
X = Stack by accident —[Δ Architecture · Δ Evals · Δ Cost · Δ Security]→ Y = Stack by decision
The hook
In 2026 almost every deck says AI. Diligence asks a harder question: is it your IP, your data and your evals — or someone else's API with a prompt on top?
Intent · Give founders and boards an honest read on architecture, AI substance, unit cost and security posture — then a sequenced plan to fix what actually threatens scale.
4
stack tiers: no-code → services
2wk
audit to written verdict
90d
remediation plan horizon
−40%
typical cost per AI task
The movement
X · today
Architecture inherited from the first prototype, AI bolted on top
Y · destination
Architecture matched to traffic, margin and diligence — with evals and cost under control
The right answer may be no-code, low-code, a modular monolith or services — and choosing the most sophisticated option is usually the most expensive mistake a pre-scale company makes.
For AI products, defensibility no longer sits in model access. It sits in proprietary data, domain evaluation sets, workflow depth and cost per task — all of which we measure rather than assume.
We expose the economics: what each user, request, tenant and inference actually costs you today, at 10× and at 100×.
Stack strategy
- No-code → low-code → modular monolith → services decision, justified by traffic and team size
- Build vs buy vs integrate priced per component, including exit cost
- Architecture scan, dependency map and single points of failure
- DORA delivery metrics: lead time, deploy frequency, failure rate, recovery
- Technical debt register priced against roadmap velocity
Applied AI, 2026
- Evaluation harness in CI: task accuracy, regression gates, human review sampling
- Retrieval quality — grounding, recall, citation and hallucination rates
- Agentic workflows scoped honestly: where autonomy pays and where it adds failure modes
- Model routing, caching, batching, small models and distillation for cost control
- Fine-tuning vs context engineering vs tools — decided on evidence, not fashion
- Data rights, provenance, tenancy isolation and training-data exposure
Security, privacy & AI governance
- Identity, secrets, PII flow and data residency review
- Prompt injection, jailbreak and abuse testing for LLM surfaces
- Audit logging and human-in-the-loop controls for regulated buyers
- SOC 2 / ISO 27001 readiness path and evidence collection
- GDPR, DPDP and EU AI Act obligations mapped to your product surface
- Vendor and sub-processor risk review
Platform economics & scale
- Cost per request, per user, per tenant and per AI task
- Cost model at 10× and 100× volume with the breaking point named
- Multi-tenancy, isolation and noisy-neighbour strategy
- Observability: traces, production evals, latency and spend dashboards
- Cloud, GPU and inference commitments negotiated against real usage
Why it matters · impact
Inference cost is the new gross margin
Token, GPU and retrieval costs sit directly in cost-of-goods. Teams that never measure cost per task discover their margin problem only when volume arrives.
Defensibility lives in data and evals
Model access is a commodity. Proprietary data, domain evaluation sets, workflow depth and switching cost are what survive the next model release.
Compliance is now a sales gate
SOC 2, ISO 27001, GDPR/DPDP and EU AI Act obligations decide enterprise deal velocity. Retrofitting them mid-deal costs quarters.
Over-engineering is the common failure
Most early platforms are three architectural steps ahead of their traffic and paying for it in cash, latency and hiring difficulty.
How it runs · deep dive
Audit · What you actually run
A two-week read of the system as built, not as described in the deck.
- Architecture and dependency scan; single points of failure named
- Stack-tier verdict: no-code, low-code, modular monolith or services — matched to traffic and team size
- Cost per request, per user, per tenant and per AI task
- Delivery health: lead time, deploy frequency, change failure rate, recovery time (DORA)
- Security posture: identity, secrets, PII flow, data residency, third-party exposure
AI substance · Evidence over adjectives
We test the AI claim the way a technical investor will.
- Evaluation harness: task-level accuracy, regression suites, human review sampling
- Retrieval quality — grounding, recall, hallucination and citation rates
- Agentic workflows scoped honestly: where autonomy pays, where it just adds failure modes
- Model routing and cost control: small models, caching, batching, distillation, fallback paths
- Data rights, provenance, tenancy isolation and training-data exposure
- Guardrails, prompt-injection and abuse testing; audit logs for regulated buyers
Platform economics & scale
The numbers that decide whether growth increases or destroys margin.
- Cost model at 10× and 100× current volume, with the breaking point identified
- Multi-tenancy, isolation and noisy-neighbour strategy
- Observability: traces, evals in production, cost and latency dashboards
- Build vs buy vs integrate priced per component, including exit cost
- Technical debt register — priced against roadmap velocity, not moralised
Remediate · A sequenced plan
Ranked by risk to revenue, not by engineering preference.
- 90-day remediation plan with owners, effort and expected cost impact
- Compliance path to SOC 2 / ISO 27001 and AI-governance documentation
- Hiring and vendor plan matched to the chosen architecture
- Architecture decision records so the next team inherits reasoning, not folklore
Measured on · KPIs
Cost per AI task
Fully-loaded inference, retrieval and orchestration cost per completed task
−40% at equal quality
Eval pass rate
Task accuracy on a domain evaluation set with regression gating
> 90%, no regressions shipped
Hallucination / ungrounded rate
Responses without valid supporting evidence
< 2%
p95 latency
95th percentile end-to-end response time
< 2s interactive
Deploy frequency
Production deployments per week
daily or better
Change failure rate
Deployments causing incident or rollback
< 15%
Uptime
Availability against the customer-facing SLO
≥ 99.9%
Infra cost per active user
Total cloud and model spend divided by monthly actives
falling quarter on quarter
Compliance readiness
Controls evidenced for SOC 2 / ISO 27001 and AI governance
audit-ready in 2 quarters
What you walk away with
- →Architecture and risk report with named single points of failure
- →Cost model at current, 10× and 100× volume
- →Evaluation harness running in CI
- →90-day remediation plan and architecture decision records
Not for
- ✕Teams wanting validation of a stack decision already made
Other programs
P1
XY Incubation
0 → Revenue, engineered.
P2
XY Acceleration
Revenue → Scale, without chaos.
P3
XY Growth Fund
Create, boot-strap, fund, scale, repeat.
P4
Global Market Access
You win on home turf, then you conquer another.
P5
XY Mentoring
One missing node, not one missing strategy.
P6
Business Model Consulting
Beyond the fluff — a calibrated engine.
P7
Product Management
Discover value, frame it, validate it, create it.
P9
Investor & Portfolio Services
Top 1% funnel. 360° evaluation.