Kenneth Teh / Lead Engineer← Portfolio

Demonstration / Engineering leverage

AI-assisted delivery: from vague request to reviewed change

An interactive Lead Engineer simulation for turning an ambiguous PM request into a codebase-grounded implementation through product discovery, API contracts, parallel AI agents, adversarial review, human approval, and iteration.

Kenneth TehWorking model

AI delivery simulator

You own the decisions. Agents produce the work.

PM request: “Can customers see why a card reward is pending?”

Confidence18%
Delivery risk62%
Elapsed0h
Rework0
1Interrogate2Prototype3Contract4Discover code5Run agents6Rival review7Human gate8Report
Stage 1 / 8Interrogate

You have room for three initial questions. Which ones reduce uncertainty before solutioning?

Latest consequence

The request has arrived. Nothing beyond the customer concern is known yet.

01

Why this workflow exists

AI makes it cheap to turn ambiguity into plausible artifacts. Plausibility is not alignment or correctness. This workflow makes uncertainty visible, assigns an accountable human at every consequential seam, and requires each stage to leave behind evidence the next stage can challenge.

02

Parallelism starts after alignment

Agents do not begin implementation from the PM's sentence. Product intent, an 80% concept, API semantics, QA cases, and repository discovery come first. Only then can frontend, backend, contract, and observability work proceed in parallel across explicit boundaries.

03

Review must be able to reverse confidence

A rival agent receives the requirement, contract, diff, and tests without inheriting the builders' narrative. Its role is to construct counterexamples and recommend rejection when the evidence is weak. A human then judges product intent, technical safety, and which findings matter.

  • A passing build does not overrule a violated product invariant.
  • Confidence should fall when a reviewer discovers a real counterexample.
  • Agent consensus never auto-approves a human gate.
  • Iteration closes findings and records what the team learned.