SOLO BUILD BY ARIEL MAGALSO · LIVE PIPELINE — NOT A SCRIPT

Verdict.

//ARIEL MAGALSO · APPLIED AI ENGINEER

I ship AI that knows when to stop

I build production AI workflows that combine LLM reasoning with evidence gates, deterministic rules, evaluation, and human approval. Verdict is the live proof: an end-to-end lead-qualification system designed and built by one person.

Run the live demo Fit it to your business

No signup. Seeded fictional data. No external data is sent.

60 labeled eval cases 11 staged pipeline steps 100% human-owned writes

//SHORT ON TIME?

A five-minute route through the engineering.

Start with one decision, then follow the artifact trail: architecture, evaluation misses, inspectable operations, and source ownership.

//BUILT WITH

Python FastAPI SQLAlchemy Claude API PostgreSQL Docker Pydantic Jinja2

//FIT VERDICT TO YOUR WORKFLOW

LOCAL PREVIEW · ZERO DATA SENT

The safeguards stay. The workflow adapts.

Choose the operating context and Verdict will map a private, illustrative integration blueprint. Change the inputs and see how the evidence, rules, approval points, and rollout would adapt.

04
Industries
04
Controls
00
Data sent
Verdict control roomIntegration blueprint
InputsWorkflowSafeguards
Live preview

//CONFIGURE THE HANDOFF

Four choices shape the preview. Change any field and the blueprint updates immediately.

Private by designNo form submission, analytics event, storage, or network request.

//ABOUT THIS PROJECT

At its core: AI that knows when not to act

The numbers below are measured from the live pipeline and its evaluation suite — each one represents a failure mode this system was built to refuse.

60

Labeled eval cases

Development and held-out sets

0

Scores on thin evidence

The gate refuses to guess

4

Responsible outcomes

Never a fifth guess

$0.01

Cost per processed lead

Haiku + cached scenarios

//THE FOUR OUTCOMES

One inbox. Four responsible outcomes.

01
Qualified

Sufficient evidence, meets the ICP — assigned, drafted, ready for a rep.

  • Evidence-cited score
  • Territory routing
  • Outreach draft for approval
02
Insufficient evidence

Can't be scored responsibly — returns the exact questions that would unblock it.

  • No number emitted
  • Unblocking questions
  • SDR verification task
03
Disqualified

Scored, doesn't meet the ICP — reason recorded, no rep time spent.

  • Explicit rule reason
  • Ruleset version recorded
  • Zero outreach sent
04
Duplicate / merge review

Matches an existing record — activity attached, never a silent duplicate.

  • Identity-first matching
  • Activity attached
  • No duplicate records

//THE THESIS

Every automation starts with the ways AI fails

An LLM asked to score leads scores everything a confident 78. Asked to research a company it doesn't know, it invents a plausible industry. Both write straight into the CRM, where fabrication becomes indistinguishable from truth. Verdict is designed backwards from those failures — every safeguard exists because the naive version visibly breaks without it.

GUIDED SCENARIO

The ambiguous lead

“Fieldwork Group” resolves to three candidate companies. Verdict declines to guess.

0

Scores emitted

3

Unblocking questions

1

SDR task created

Run this scenario

//THE PIPELINE

Judgment work, staged and gated

//01

Resolve & validate

Normalize the submission, resolve contact and company identity, and check for duplicates before any model spend.

//02

Research with citations

Enrich company facts against approved sources. Every fact carries a source link and a verified / uncertain / conflicting status.

//03

Gate, then score

An evidence-sufficiency gate runs first. Past the gate, deterministic ICP rules produce an explainable score — the model never picks the number.

//04

Approve & write

Outreach drafts and CRM changes are proposed as diffs. A human approves; writes are idempotent and fully audited.

Review the full architecture

//THE DIFFERENCE

Naive AI qualification vs Verdict

LLM straight into the CRM

Scores every lead a confident 78
Invents company facts under uncertainty
Writes fabrications straight into the CRM
Creates a duplicate for every repeat email
Sends outreach nobody reviewed
No trace of why a decision happened

Verdict.

Refuses to score below the evidence floor
Marks unknowns as unknown, with questions
Proposes CRM changes as reviewable diffs
Resolves identity before creating anything
First outreach always requires approval
Every stage lands in an audit trail

//THE MIND BEHIND THE WORK

Ariel Magalso

Applied AI Engineer · Philippines · Remote-ready

I independently designed and implemented Verdict's pipeline architecture, application, evaluation suite, database, operational monitoring, interface, and deployment. I turn ambiguous business workflows into reliable, human-in-the-loop systems where traceability matters as much as speed — and I can show the code, test, and operating signal behind each claim.

  • AI workflow architecture
  • CRM and operations automation
  • Structured LLM extraction
  • Evaluation and reliability testing
  • Human approval workflows
  • Monitoring, cost controls, and audit trails

//HIRING?

Looking for someone who ships AI you can trust

Available for AI automation roles and consulting — workflow design, LLM evaluation, and production automation with real safeguards.

Read the case study Email Ariel LinkedIn Full portfolio

//PROOF LOG

Milestones you can verify

AUG 2026 Recruiter-ready Python port shipped Live demo, evals, operations telemetry
AUG 2026 60-case evaluation suite Zero false scores on insufficient-evidence cases
AUG 2026 Prompt-injection scenario shipped Untrusted page content flagged, never obeyed
AUG 2026 Durable worker deployed Spend caps, rate limits, cached guided scenarios
Read the case study Review 60-case evals Inspect operations Read the source

//ENGINEERING PROOF

Every important claim has a place to inspect.

The portfolio is designed as an evidence trail: run the behavior, inspect the boundary, measure the failure modes, and read the implementation.

01

Behavior

The guided demo makes each pipeline stage visible and explains why the final outcome happened.

Run a scenario →
02

Measurement

The evals page separates development from held-out cases and shows false scores, false refusals, and confidence intervals.

View the scorecard →
03

Operation

The operations page reads persisted audit-event data for latency, spend, completion, stuck work, and duplicate prevention.

Inspect the signals →

//THE PORTFOLIO STATEMENT

“The AI meant to clean your pipeline shouldn't become the fastest way to poison your system of record.”

THE FAILURE THESIS VERDICT IS BUILT AGAINST

//WHY IT HOLDS UP

Four properties, enforced everywhere

Evidence-linked

Every company fact and score criterion cites a source a human can open and check.

Deterministic where it counts

The model extracts and classifies; plain code applies the rules and arithmetic.

Human-in-the-loop

Consequential actions — outreach, CRM writes, merges — wait for explicit approval.

Operable

Latency, cost, failures, and audit events are visible on an inspectable operations page.

//ILLUSTRATIVE CALCULATOR

Measure the tradeoff, not the hype.

Adjust the assumptions for your own operation. These are not measured customer results — this demonstrates the calculation, seeded with the fictional example from the product plan (800 leads/month, 12 minutes of manual work per lead, 65% auto-eligible).

LEADS ELIGIBLE FOR AUTOMATIC HANDLING

520

per month

STAFF HOURS POTENTIALLY RETURNED

104

per month

STILL ROUTED TO HUMAN REVIEW

280

per month, by design — not a shortfall

ESTIMATED NET MONTHLY SAVINGS

$3,632

$3,640 handling-cost reduction − $8 automation cost

//FAQS

Before you run the demo

Run it now
Is this a real working system or a mockup?

A live pipeline. The demo runs real identity resolution, research, gating, and scoring against seeded fictional data. Guided scenarios are idempotent on their submission id, so repeat clicks return the already-completed result instantly and for free.

Why does it sometimes refuse to give a score?

A number produced on thin evidence is noise that erodes rep trust. If fewer than the configured floor of ICP criteria can be resolved with evidence, Verdict emits no score and instead returns the specific questions that would unblock scoring.

What stops the AI from inventing company facts?

Facts only enter the system with a source URL and verification status, and outreach drafts are checked against those sources. Unsupported claims are stripped before a human ever approves a message.

What happens if a researched page tries to manipulate the AI?

Page content is treated as untrusted data. One guided scenario plants a prompt-injection payload in a seeded source — the instruction is ignored, flagged in the audit log, and never reaches a draft.

Can I see the code and the evaluation results?

Yes — the source is on GitHub, and the evals page publishes the full scorecard, including both failure directions: false scores and false refusals.

//AI SAFEGUARDS SERIES

Three systems, one argument

Each one automates judgment work an LLM will confidently get wrong — and each is built against a bigger consequence than the last. The safeguards get harder as the cost of being wrong goes up.