AI automation portfolio project

Support answers thatshow their work.

A citation-first AI support agent for product compatibility. It retrieves evidence, measures confidence, and hands uncertainty to a human—before an unsupported answer reaches a customer.

Test the agent See how it works

Source-linkedEvery supported answer

Evidence-gatedNo confident guessing

Human-readyClear escalation path

ANSWER QUALITYDEMO

92%

Evidence confidence

Sources attached3 verified records

✓

Evidence gate passedResponse safe to send

Interactive prototype

Put the evidence gate to work.

Ask a compatibility, accessory, or warranty question. Try an unknown product to see the human-review path.

✓Evidence gate passed

DETERMINISTIC DEMO

GROUNDED ANSWER

No. The DLC-9 is a narrow detail blade made for the Vectra T9 trimmer interface. The Vectra V2 clipper uses the wider F45 blade mount, so the two parts should not be combined.

Evidence confidence96%

Calculated from retrieval strength and agreement—not the model's opinion.

DECISION TRACE

1Request validated
22 matching records retrieved
3Evidence gate passed

SUPPORTING SOURCES2 records

[1] kb-blade-dlc9

BL-DLC9 Detail Blade — Compatibility Matrix

Compatibility / Blades

“Approved only for the Vectra T9 trimmer mount; not compatible with clipper mounts.”

[2] kb-vectra-v2-clipper

Vectra V2 Clipper — Product Reference

Catalog / Clippers

“The V2 cutting head accepts F45-series clipper blades.”

Portfolio demonstration using a fictional grooming catalog and synthetic evaluation data. No real customer or business results are claimed.

System architecture

One question. Six controlled decisions.

The language model is one step in the workflow—not the source of truth. Retrieval, thresholds, citations, and escalation are controlled in application code.

Question intake

Validates the request and strips unsafe or empty input.

Local retrieval

Ranks product, policy, and compatibility records by relevance.

Evidence gate

Blocks an answer when the supporting evidence is too weak.

Grounded response

GPT-5.6 can compose from approved evidence only; demo mode works offline.

Citation builder

Attaches the exact records used so every claim can be checked.

06done

Human review

Routes uncertainty and unsupported requests to a person instead of guessing.

Grounded by design

The assistant can answer only from the curated knowledge base.

Human where it matters

Ambiguous compatibility and warranty cases are escalated.

Observable workflow

Confidence, evidence, mode, and decision steps stay visible.

Demo evaluation

Measured for reliability, not theatre.

Automated fixtures check known compatibility, conflicting terminology, citation attachment, and out-of-catalog requests on every build.

Metrics below describe the included synthetic test set. They are reproducible in the repository and are not presented as client outcomes.

Grounded test answers100%Sources attached when evidence exists

Escalation scenarios4/4Routed safely

Deployment modes2Deterministic + OpenAI

Build checks activeRetrieval · API · render

What this demonstrates

Production-shaped thinking in a focused prototype.

AI integration

Server-side Responses API integration with a no-key deterministic fallback.

Workflow design

Explicit evidence thresholds, refusal behavior, and human-review routing.

Full-stack delivery

Responsive UI, typed API contract, test fixtures, documentation, and deployment.

AI automation & integration

Have a workflow that needs fewer guesses?

I design focused AI systems that connect knowledge, APIs, and human approvals into a workflow your team can inspect and control.

Review the live demo