Skip to content
All case studies

Enterprise AI delivery

Trading and advisory firm

The 10-week engagement that paid for the entire platform.

A trading and advisory firm hired a single director and an agent roster. In 10 weeks they shipped 8 distinct projects, including a 40,000+ line production Regulatory Intelligence Platform with hybrid RAG, three-tier model orchestration, and WORM-compliant audit trails. The work product is real. The audit trail is reviewable.

Single director. 9 specialist agents. 10 weeks.

10
Week
Engagement
8
Projects
Delivered
40K+
LOC Production
Python
9
Domain Agents
Deployed
40
Regulatory
Documents
389
Captured
Sessions
Human-team baseline

Tech Lead + 2 Senior Backend Engineers + Data Engineer + AI/ML Engineer + Regulatory Analyst + Compliance Officer + QA Engineer + Project Manager.

8-person consulting team · 4–6 months elapsed · ~$300,000–$500,000 all-in cost equivalent.

Agentic delivery

Single director. 9 specialist agents (regulatory expert · portfolio proxy · risk manager · compliance/legal · AI governance · data engineer · technical architect · QA validator · project manager) deployed as code.

~250+ director-hours over 10 weeks · API + tooling $2,000–$5,000 for the entire engagement.

What a Head of Engineering will ask in a first call. Pre-answered.

Detailed audit trails, methodology, and source code are reviewable by qualified evaluators under NDA. The headlines are below.

Q1
How many of the 40K+ LOC are agent-generated vs human-written?

Detailed human/agent split is reviewable under NDA with the engagement audit trail. Headline: every commit went through HITL gates (Triage Accept → Plan Approval → Verification Accept) before merge. The gating is non-skippable - agent-authored code that did not pass review did not ship. Defect-rate-on-agent-generated and review-time-per-shift figures are part of the same audit trail.

Q2
Which agents shipped which projects?

See "Project-by-project breakdown" below. Each project lists the specialist agents that contributed. No advisor-only / no-output agents are listed - the roster reflects builders who shipped code or schemas under HITL approval. Agent definitions are YAML files in vault - diff-able and rollback-able.

Q3
What is the human-supervision ratio?

Single director + agent teams across 5+ specialist roles per project. Approximately 250+ director-hours over the 10-week engagement (HITL approval, brief authoring, edge-case escalation, final review). The director was not writing the code - they were directing the agents. API + tooling cost in the $2,000–$5,000 range for the entire engagement.

Q4
How is this auditable / verifiable?

Three layers. (1) Git diff URLs for every project - agent-authored commits carry agent attribution metadata. (2) Vault session telemetry: 389 captured operational sessions including agent prompts, gate decisions, and review notes. (3) Methodology: every project records human-team baseline at task close and writes to an efficiency database - the multiplier numbers are computed, not estimated. Available to qualified evaluators under NDA.

Q5
Why is this not just "AI-assisted coding" by another name?

AI-assisted coding accelerates an individual engineer at the keystroke level. Plexi Flexor staffs specialist roles. Across the 10-week engagement, 9 distinct agent identities (regulatory expert, portfolio proxy, risk manager, AI governance, data engineer, technical architect, QA validator, project manager, compliance/legal) operated as the team - each with its own vault context, voice, and approval surface. The director coordinated specialists; the specialists were agents. The closest equivalent is hiring a 6–8 person consulting team, except they live in YAML.

8 projects. Which agents shipped what.

P1

Regulatory Intelligence Platform

4–6 wk ~40,000 LOC

Production-grade regulatory intelligence platform with hybrid RAG (dense embeddings + sparse retrieval + reranking), three-tier model orchestration, and WORM-compliant audit trails (7-year retention).

Output: Python · ~10 microservices · 6 data-source connectors · state-machine orchestration · HITL gates on material-exposure decisions.

Specialist agents: regulatory-expertcompliance-legaldata-engineerqa-validator
P2

Domain Knowledge Base

2–3 wk

Curated regulatory document corpus across multiple jurisdictions with overlap reasoning, jurisdiction conflicts surfaced, and citation provenance preserved.

Output: ~40 curated regulatory documents · 5 market verticals · jurisdiction-conflict ontology.

Specialist agents: regulatory-expertcompliance-legalarchivarius
P3

Specialist Agent Roster

2 wk

Curated set of specialist agents tuned for trading & advisory operations: regulatory expert, portfolio proxy, risk manager, compliance/legal, AI governance, data engineer, technical architect, QA validator, project manager.

Output: 9 domain-specialized agents · YAML-defined · vault-versioned · simulation-harness validated before production.

Specialist agents: ai-ml-architectproject-managerqa-validator
P4

Network Intelligence Platform

1–2 wk

Counterparty graph + network intelligence over the firm's active engagements: who knows whom, what conversations are warm, where adjacency creates opportunity.

Output: Graph schema · ingestion from CRM + comms · weekly cockpit with deal-velocity flags.

Specialist agents: data-engineeranalyst
P5

Integrated Conversational Platform

1–2 wk

Authenticated conversational interface for portfolio queries, risk introspection, and regulatory lookups - backed by the firm's data substrate, not a third-party LLM with raw fund data.

Output: Web UI + API · session-scoped auth · all queries audit-logged · grounded retrieval.

Specialist agents: ai-ml-architectdata-engineer
P6–P8

Horizon-1 / Horizon-2 platform builds + 19+ initiative specifications

~3 wk

Horizon-1 platform components shipped to production. Horizon-2 designs documented with PRDs, schemas, and cost models. Plus 19+ documented initiative specifications across compliance monitoring, AML-KYC automation, edge analysis, capacity estimation, and conversational investment intelligence.

Output: 19+ initiative specs · production cutover plan · 100 events / 82 task completions / 9 repos active in first month.

Specialist agents: project-managerbusiness-analystsolutions-architect

Don't take the numbers - verify them.

What separates this case study from a competitor brochure: the audit trail exists, and qualified evaluators can read it. We will walk you through commit-by-commit on the technical call.

Git diff URLs

Every agent-authored commit carries agent attribution and HITL gate timestamps. Buyer-grade verification: pick a commit, read the diff, see which agent authored, see which director approved, see which gate captured.

389 captured operational sessions

Session-level telemetry including prompts, gate decisions, review notes, and rejection reasons. The audit trail is complete; redacted exports available under NDA.

Self-rewrote dogfood loop (Plexi Flexor site)

plexiflexor.ai itself was rewritten by the platform it sells. April 27, 2026: an internal integrity-auditor advisor flagged 24 issues across the portfolio. April 29: 12 had shipped to UAT under HITL approval. The diffs are in the repo. Same loop. Same gates. Same agent roster.

~40 curated regulatory documents · 5 market verticals

Domain Knowledge Base content with citation provenance, jurisdiction-conflict resolution, and overlap reasoning. Available for qualified evaluator inspection under NDA.

All case studies

Updated September 14, 2026