For teams shipping AI to production

AI that passes review before it reaches your customers.

Evals, guardrails, and agent systems built to hold up in front of auditors, customers, and regulators. Delivered by a cleared engineer who evaluates models for frontier AI labs and has shipped these systems in production.

Book a 30-minute risk reviewSee a sample audit report
Trusted by frontier AI labsOpenAIMetaGoogle
4/6
Assurance scorecard · sample
Support agent, pre-launch
  • Faithfulness and hallucinationPASS
  • Prompt injection and SSRFPASS
  • Human approval gatesPASS
  • Spend and rate controlsREVIEW
  • Audit loggingPASS
  • Data access scopingREVIEW

Every engagement ends with a scorecard like this, mapped to NIST AI RMF.

94%
cut in clinical content validation time at Mayo Clinic
100s
OpenAI model deliverables scored on calibrated writing-quality evals
27
named AI failure patterns in four categories, caught by unslopify
10+
years shipping production software
Trusted by
  • OpenAI (via Mercor)
  • Meta (via Handshake AI)
  • Google (via Mercor)
  • Mayo Clinic
  • Signet Jewelers
  • Polaris I/O
  • JP Morgan Chase
  • CACI / US Army
  • Veterans Affairs
  • Bayer
  • Rivian
  • College Board
  • Live Nation
  • The Infatuation

Services

Three ways to make your AI safe to ship.

01 · Fixed fee · 2 weeks

AI Risk Audit

An outside read on the system you already have, before a launch, a board review, or a procurement questionnaire.

  • Threat model of prompts, tools, and data paths
  • Eval gap analysis against your real failure modes
  • NIST AI RMF and EU AI Act mapping
  • Written report and scorecard your team can act on
Ask about scope →
02 · Project · Scoped to your system

Evals and Guardrails

The test suite your model never had. Built to run in CI so regressions fail the build, not the customer.

  • Calibrated LLM-as-judge rubrics with written rationales
  • Red-team and prompt-injection test sets
  • Deterministic gates that need no model call
  • Retrieval quality measured with RAGAS
Ask about scope →
03 · Build · Scoped to your system

Safe Agent Delivery

Agentic systems designed with controls from day one, so a human signs off where it matters.

  • Human approval gates at irreversible steps
  • Scoped tool permissions and SSRF protection
  • Per-agent spend and rate controls
  • Audit logs your compliance team can read
Ask about scope →

Method

How every engagement runs.

  1. 1

    Threat model

    Map who uses the system, what it can touch, and what failure costs.

    Risk register
  2. 2

    Evals

    Turn those failures into tests with rubrics a reviewer can audit.

    Eval suite in CI
  3. 3

    Guardrails

    Add gates, permissions, and human sign-off where tests are not enough.

    Controls + scorecard
  4. 4

    Monitor

    Track drift, cost, and reviewer feedback after launch.

    Alerts + reports

Production findings feed the next threat model

Work

Risk, control, result.

Filter by industry or kind of work. Every engagement opens to the full story, and the AI ones include an explainer at three levels of detail.

Frontier AI labs

Evaluation and expert data
  • OpenAI, via MercorOpenAIWriting-quality and presentation evals

    Expert evaluation on OpenAI's Feather platform: catching low-substance AI writing in model-generated business documents and slide decks, each score on a calibrated 1 to 7 rubric with a written rationale.

    Hundreds of OpenAI model deliverables scored

  • Meta, via Handshake AIMetaCoding-agent evaluation

    Hard coding tasks with hidden verifiers, then side-by-side agent runs ranked on independent evidence rather than self-reported success.

    Hard tasks, hidden verifiers, evidence-backed rankings

  • Google, via MercorGoogleExpert writing for model training

    Expert writing for Google through Mercor, producing the human-written work Google's models learn from.

    Selected as an expert writer

More engagements13

Open tools we build in public

unslopify

A deterministic linter for AI-written prose. 27 named patterns, zero model calls, CI-ready exit codes. MIT licensed on PyPI.

Grasp

Plain-language briefs that explain merged code changes to every reader, from new engineer to executive.

See what an audit actually delivers.

A redacted sample report: threat model, eval results, scorecard, and the fix list, exactly as clients receive it.

Request the sample report

About

Engineering depth, policy fluency.

Fareez Ahmed founded V2 Software after a decade of shipping production systems for healthcare, finance, retail, and defense. He builds the evals and controls himself, and he has worked inside the review processes they have to survive.

His MA in International Affairs and Economics from Columbia means the regulatory side of AI is familiar ground, not an afterthought. He holds an active Secret clearance and works remotely from San Francisco and Dallas.

Active Secret clearanceDefense and federal ready
MA, Columbia UniversityInternational Affairs and Economics
Four languagesEnglish, French, Arabic, Bengali

Shipping AI into a high-stakes setting?

Book a 30-minute risk review. You leave with the three risks worth fixing first, whether or not we work together.

Book a 30-minute risk review
fareez@v2software.app(202) 677-0160San Francisco, CA · Dallas, TX · RemoteLinkedIn