Skip to content
Services ·04 · Technical

Diligence that reads the repo.

Architecture review, model selection, evaluation, and technical due diligence — led by Valentin Badonfai, who has built and shipped production AI systems end-to-end. Not a consultant who's read about it.

Engagement
Scoped or retainer
Lead
Valentin Badonfai
Timeline
2–6 weeks scoped · ongoing optional
Best fit
Builders & acquirers
Deliverable · technical DD report
structure used for acquirer-side and growth-equity engagements
SECTION 03 · MODEL STACK
Production model selection & routing

3.1 — Model inventory: 3 production paths (GPT-4o, Claude 3.5 Sonnet, internal fine-tune), routed via custom layer.

3.2 — Routing logic: read in src/router/index.ts. Keyword + intent match, fallback chain configured. Sound for current scale.

3.3 — Cost: $0.018/req median, $0.043 p95. Fine-tune use at 14% of volume. Fully GPU-amortised at >800k req/day.

3.4 — Risks: single-vendor dependency on Anthropic for high-trust workflows. Sub-recommendation §6.2.

3.5 — Eval coverage: 61% of high-stakes paths covered. Below acquirer benchmark of 80%. See §5.

p. 14 of 38 amafi/dd · Q1 2026
Table of contents · what's inside
01
Executive read
One-page TL;DR for the IC. Verdict, top three risks, top three strengths.
2pp
02
Architecture & system design
Diagrams, data flow, deployment topology, scaling assumptions.
6pp
03
Model stack & routing
Provider mix, fine-tune use, fallback chains, vendor dependency risk.
5pp
04
Cost & latency model
Per-token economics, latency p50/p95/p99, scenarios for model substitution.
4pp
05
Eval framework audit
Coverage, methodology, drift detection, gap list with severity.
5pp
06
Risk register
Named risks, ranked by severity × timing, with recommended mitigations.
3pp
07
Engineering quality
Code quality signal, test coverage, CI/CD maturity, hiring bench.
4pp
08
Recommendations
Pre-close conditions, 90-day plan, year-one engineering investments.
3pp
Output of Mode 03 · Transaction-side DD Sits next to the financial DD binder · same rigour, same language

Three modes. One opinion.

Architecture and eval review for builders. Cost and latency audit for CFOs. Transaction-side DD for acquirers. Each scoped tightly, each ending in a written opinion.

01
Architecture & eval review
Read the system end-to-end. Stress-test the eval harness. Give an opinion on model choice, routing, retrieval, and what the actual moat is — versus what the deck claims.
2–4 weeks
02
Cost / latency audit
Per-token economics, retrieval overhead, response latency under load, model substitution scenarios. The numbers a CFO and a CTO will both ask for.
1–2 weeks
03
Transaction-side DD
Technical DD for acquirers and growth-equity LPs. Output a report that sits next to the financial DD binder — same rigor, same language.
scoped

Deliverables, named.

01
Architecture opinion
A written read on the system — model choice, routing logic, retrieval, eval harness, and the failure modes. With recommendations, not just observations.
02
Cost & latency model
Per-token economics by use case, latency distributions under load, and substitution scenarios for model swaps. A live spreadsheet, not a static report.
03
Eval framework audit
Honest read on whether your evals actually measure what matters, what they miss, and what an acquirer's diligence team will pick at.
04
DD report (acquirer)
A 20–40 page technical DD report for the acquirer's binder. Models, data, eval, infra, security, hiring quality. Sits next to financial DD.
05
Risk register
Named technical risks ranked by severity and timing. The page the deal team takes to the IC meeting.
06
Engineering sourcing
Optional: read on the engineering bench, hiring plan, and the senior gaps that will show up in year two if not addressed pre-close.

Want a real technical read?

Thirty minutes, confidential, no deck required. Tell us what you're building or buying — Valentin will tell you honestly what's worth a deeper look.

Book a confidential intro → Pick a slot →