Miguel Almeida
← Selected work

Read-only code intelligence

Atlas

Deterministic risk first. Model explanation second.

Code-intelligence tools lose trust when a model invents the underlying risk or tenant context leaks across analysis boundaries.

Architecture case study. Public narrative is limited to verified system boundaries and known limitations.

Read case studyPrivate repositoryNo public demo
RoleProduct architecture · frontend · trust boundaries
ContextIndependent system
StackTypeScript · PostgreSQL · RLS · retrieval · audit trails

Ownership

What I built and why it mattered

Designed the product around deterministic risk generation, request-scoped database context, explicit workspace and ingestion-run scope, and retrieval used only to explain established findings.

PostgreSQL RLS · request-scoped context · workspace → repository → ingestion-run scope

Interactive boundary

The model explains a risk it did not invent.

Deterministic system

Risk exists before model language.

Workspace, repository, and ingestion-run context produce the finding and its evidence.

  1. 01Scoped request context
  2. 02Deterministic rule
  3. 03Evidence location
  4. 04Audit record

System flow

The working sequence

  1. 01Repository ingestion
  2. 02Deterministic risk generation
  3. 03Evidence retrieval
  4. 04Model-assisted explanation
  5. 05Audit review

Technical judgment

Decisions that define the architecture

01

Generate risk deterministically before any model explanation.

02

Scope rows by workspace, repository, and ingestion run.

03

Set database identity for each request and enforce it through RLS.

04

Retain audit trails around ingestion and explanation.

Evidence ledger

Claims with their boundaries attached

Verified

Deterministic risk boundary

The underlying risk is generated before model-assisted explanation.

Atlas architecture notes
Verified

Tenant isolation

Workspace, repository, and ingestion-run scope is enforced through request context and PostgreSQL RLS.

Atlas architecture notes

Current limits

What this work does not establish

  • No production adoption claim is made.
  • Model explanations still require review.
  • Performance observations are only useful with their dataset and environment context.

Need a shorter path?

Ask MiguelLLM for the engineering version.