Othmane Achir achirothmane

Verification & control infrastructure for AI, automation, and high-consequence software

I build systems that decide when there is enough evidence to act — and when the correct answer is UNKNOWN.

What I build

I build verification and control systems for software that should not act on weak evidence.

My current work focuses on:

AI-agent control · CI reliability · causal verification · developer tooling · high-consequence automation

I work primarily with:

Python · GitHub Actions · PostgreSQL · CI/CD · AI/LLM systems

The recurring engineering question behind my projects is:

What evidence must be established before this system is allowed to act?

Open to

Verification / reliability / developer-infrastructure work, applied AI systems, and selected technical collaborations where evidence, control, and auditability matter.

Most relevant systems

CI Retry Gate

Problem
CI systems often retry failures without knowing whether retrying is actually justified.

Built
An evidence-gated decision layer that checks provenance, causal evidence, side effects, and retry limits before granting rerun authority.

Why it matters
Fewer blind reruns, clearer failure handling, and auditable retry decisions.

failure → evidence → decision → retry / block

Consequence Boundary Completeness

Problem
An AI agent may reach a real-world consequence through a path that bypasses the approval boundary intended to control it.

Built
A scanner and runtime witness for modeled consequence paths, expected boundaries, bypasses, and unresolved evidence.

Why it matters
Authorization controls are only useful if alternate paths cannot silently route around them.

agent → path analysis → boundary → counterexample / covered / unknown

PostgreSQL Change Safety

Problem
A workload regression after a PostgreSQL change is easy to observe and easy to misattribute.

Built
Comparable workload windows, stable SQL fingerprints, controlled causal experiments, plan-variant analysis, and explicit confounder handling.

Why it matters
The tool separates “a regression happened” from “we have enough evidence to say why.”

before/after → regression → causal isolation → evidence strength

Legal Authority Diff

Problem
A legal-AI answer can still look plausible while its citations, authority strength, treatment, or claim support become weaker.

Built
Differential regression testing across baseline and candidate outputs, with explicit UNKNOWN and WORLD_CHANGE states.

Why it matters
A citation existing is not the same as the citation being sufficient support for the claim.

baseline → candidate → authority diff → block / unchanged / world change

Also building

Private Code Modernization Factory — evidence-first modernization analysis with constrained patch proposals, differential verification, and escalation when automation has not earned authority.

Technical focus

The engineering thesis

A recurring failure pattern appears across AI agents, CI systems, databases, legal AI, and software automation:

Plausibility is not authority. A system should act only when the evidence required for that action has actually been established.

The architecture I keep returning to is:

Input / Event / Proposed Action
              │
              ▼
      Evidence Collection
              │
              ▼
         Verification
        ┌─────┴─────┐
        │           │
   sufficient   insufficient
        │           │
        ▼           ▼
Authority/Policy   UNKNOWN
        │        BLOCK / ESCALATE
        ▼
    Decision Gate
     ┌───┼───┐
     ▼   ▼   ▼
   ALLOW BLOCK HUMAN
        │
        ▼
     Execution
        │
        ▼
 Outcome Verification
        │
        ▼
   Audit / Replay

Design rules I care about

Explicit uncertainty `UNKNOWN` is a valid engineering outcome. It is safer than inventing certainty from weak evidence.	Fail closed High-consequence actions do not silently inherit permission when evidence is incomplete.	Read-only first Observe and prove value before enabling mutation, reruns, quarantine, merge, deploy, or other write authority.
Differential verification Prefer controlled before/after evidence over plausible explanations.	Causal isolation When attribution matters, test competing explanations instead of treating correlation as cause.	Audit & replay Important decisions should retain enough evidence to inspect and reproduce how authority was granted.

Connect

Open to selected technical collaborations in verification, reliability, developer infrastructure, and applied AI systems.

---

Evidence before action.

Build → test → falsify → strengthen the evidence → automate only what has earned authority.

_{Independent tools first. Shared platform only when repeated real-world usage proves the same primitives belong together.}

Navigation Menu