Build the platform. Control the agents. Defend the architecture.
A production engineering programme for lead engineers, architects and senior data and ML engineers building enterprise AI platforms inside hedge funds, asset managers and investment banks. The centre of gravity is distributed systems, reusable platform capabilities, agent control, evaluation, security, observability and model risk, not prompt engineering.
Advanced practitioner to technical lead · Python first · concept brief, architecture clinic, implementation lab, failure exercise, design defence
The cohort builds one platform throughout the programme, serving portfolio managers, quantitative researchers, fundamental analysts, risk, compliance and platform engineers. Every module extends the same system rather than starting a new exercise.
Research ingestion across filings, transcripts, internal notes and structured market data; point-in-time, permission-aware retrieval with citations; issuer and sector research agents; portfolio exposure and scenario tools; investment-memo assembly with contradiction and evidence checks; research-code generation in isolated execution environments.
Multi-agent orchestration with budgets, approvals and durable state; offline and online evaluation, tracing, cost and latency controls; Kubernetes deployment, GitOps promotion and production incident response.
The programme does not claim that an autonomous agent should make unsupervised investment decisions. Portfolio decisions, trade instructions and external communications remain subject to explicit human authority, policy and auditable controls. Aegis is a fictional firm created for teaching.
The programme treats the platform as a set of reusable services rather than one application, so each layer can be evolved, tested and governed independently.
| Layer | Responsibilities | Candidate technologies |
|---|---|---|
| Experience | Analyst workspace, APIs, streaming responses, approvals | FastAPI, Streamlit or React |
| Agent runtime | State machines, routing, tool execution, durable workflows | LangGraph/LangChain, LlamaIndex, custom Python |
| Model gateway | Provider abstraction, routing, quotas, caching, policy | LiteLLM-compatible gateway or custom service |
| Knowledge | Parsing, indexing, hybrid retrieval, reranking, provenance | PostgreSQL/pgvector or Qdrant, object storage |
| Data | Market, fundamental, portfolio and research pipelines | SQL, Polars, Pandas, dbt, Airflow or Dagster |
| Evaluation | Datasets, judges, deterministic checks, experiment lineage | MLflow plus a custom evaluation harness |
| Platform | Containers, orchestration, deployment and secrets | Docker, Kubernetes, Helm, Flux |
| Operations | Logs, metrics, traces, alerts, SLOs and cost attribution | OpenTelemetry, Prometheus, Grafana, Loki/Tempo |
| Governance | Entitlements, audit, data lineage, model risk, change control | OPA-style policy, IAM, evidence registers |
Each module carries its engineering labs and a design review, because the test of platform work is whether the decisions survive challenge.
Labs Map an issuer-research workflow; write the platform charter and initial ADR set.
Design review Defend scope, control boundaries, SLOs and expected value.
Labs Build a typed model gateway; benchmark three routing policies under failure and load.
Design review Approve the model abstraction and failure semantics.
Labs Create a point-in-time issuer data mart; build a tested document-to-object-store pipeline.
Design review Demonstrate that the dataset cannot leak future information.
Labs Implement hybrid financial-document retrieval; build adversarial retrieval tests.
Design review Compare retrieval designs using recall, precision, attribution and latency.
Labs Build an issuer-analysis agent with six typed tools; safely execute generated analytics code.
Design review Threat-model every tool and approve authority boundaries.
Labs Build research, risk and critic agents; implement bounded parallel orchestration with recovery.
Design review Prove that multi-agent complexity beats a simpler baseline.
Labs Implement an evidence-linked investment memo; build a portfolio scenario agent with deterministic calculations.
Design review PM and risk personas challenge the workflow’s usefulness and controls.
Labs Build the Aegis evaluation harness; run a controlled prompt/model/retrieval comparison.
Design review Release or reject a candidate using a documented quality gate.
Labs Attack and defend a retrieval agent; implement entitlement-aware tools and immutable audit events.
Design review Security and model-risk go/no-go review.
Labs Convert the prototype into distributed services; conduct load, saturation and recovery tests.
Design review Capacity and resilience review against explicit SLOs.
Labs Package Aegis with Helm; promote it through dev, test and production-like clusters using Flux.
Design review Production-readiness and rollback demonstration.
Labs Build operational dashboards and alerts; run a corrupted-index/failed-model incident game day.
Design review Executive launch decision and 12-month platform roadmap.
A portfolio manager requests a time-bounded investment review of a listed issuer. The system assembles authorised evidence, calculates portfolio exposure, evaluates a downside scenario, identifies contradictions, drafts a cited investment memo and routes the result for human approval. It must recover from partial service failure without losing auditability.
| Business value and workflow fit | 10% |
|---|---|
| Architecture and platform reuse | 15% |
| Retrieval and data integrity | 15% |
| Agent and tool engineering | 15% |
| Evaluation quality | 15% |
| Security and governance | 10% |
| Reliability and operations | 10% |
| Deployment and developer experience | 5% |
| Technical defence | 5% |
Pass standard: 75% overall, with no score below 60% in retrieval and data integrity, security and governance, or reliability and operations.
Early modules run under Docker Compose; later modules move to a local Kubernetes cluster, so the deployment work is done against a real orchestrator rather than described.
| Feature | Self-paced | Cohort | Enterprise |
|---|---|---|---|
| Format | Written chapters, video explainers and podcasts | Everything in self-paced, plus scheduled live sessions | Everything in cohort, delivered privately to your team |
| Live sessions | None | Scheduled, instructor-led | Scheduled, instructor-led, private |
| Mentorship | Not offered | Not offered | Not offered |
| Access | Lifetime | Lifetime | Lifetime for every enrolled seat |
| Pace | Entirely your own | Guided schedule with a peer group | Agreed with your desk |
| Tailoring | Fixed curriculum | Fixed curriculum | Sequenced to your markets, systems and governance |
| Best for | Individuals learning around a job | Individuals who want structure and deadlines | Desks building the same capability together |
The full chapter list, what each module covers, and how the tiers compare - sent to your inbox as a PDF.
Private delivery for your desk, sequenced to your markets and systems. Tell us the team and we will scope it.
Run this for my deskTake it as a cohort, or bring it to your engineering team as a private programme.