# How Forward Deployed Engineering is done at Kepler — Vinoo Ganesh

Source: [AI Engineer](https://www.youtube.com/watch?v=1OMHGsUZiqA)  
Feed7 permalink: https://feed7.dev/p/how-forward-deployed-engineering-is-done-at-kepler-vinoo-ganesh-0thtvuc  
Published: 2026-07-28T16:00:00.000Z  
Trust: Source Linked (source_linked)

## Why Included

Kepler frames forward deployment as product discovery: observe real work, ship the smallest useful fix, then turn repeated pain and customer vocabulary into durable product leverage.

## Source Summary

A customer requested a **47-page specification**, 14 metrics, and a three-month BI project. On-site observation revealed the immediate need was one late-truck alert, built in **4 hours**; another pipeline reportedly fell from 17 hours to about 2.

## Practical Implication

Builders should watch users perform the job, ask what happens next, and solve a small repeated pain before expanding scope. Treat copied data, tab switching, recurring tasks, and overloaded terms as evidence for tools, integrations, and a product ontology.

## Agent-Ready Context

A customer requested a **47-page specification**, 14 metrics, and a three-month BI project. On-site observation revealed the immediate need was one late-truck alert, built in **4 hours**; another pipeline reportedly fell from 17 hours to about 2.

Builders should watch users perform the job, ask what happens next, and solve a small repeated pain before expanding scope. Treat copied data, tab switching, recurring tasks, and overloaded terms as evidence for tools, integrations, and a product ontology.

Small fixes are rarely temporary: one improvised retention script spread across a nearly **100,000-person customer** and remained in use a year later. Fast delivery therefore still needs production assumptions, ownership, and a path into the core product.

## Connected Context

Feed7 judgment across 262 accumulated Signals:

This makes field observation the mechanism for finding the minimum useful deployment, not merely a discovery ideal: a large requested project can collapse into one production-worthy alert. It reinforces prior advice to encode real workflows and resist one-offs, while adding that even four-hour fixes need ownership and a route into the product because local tools can spread unexpectedly.

- [AI tools for Forward Deployed Engineering — Vasuman Moza, Varick Agents](https://feed7.dev/p/ai-tools-for-forward-deployed-engineering-vasuman-moza-varick-agents-12kjg79) — Kepler supplies a concrete method for Varick’s workflow-capture principle: observe the job, follow what happens next, and automate the repeated pain actually encountered.
- [How Forward Deployed Engineering is done at Decagon — Sunny Rekhi](https://feed7.dev/p/how-forward-deployed-engineering-is-done-at-decagon-sunny-rekhi-02myrh1) — The rapidly spreading retention script illustrates why Decagon warns against brittle one-offs and why repeated customer needs should migrate into product features.
- [Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution](https://feed7.dev/p/2607-13034v1-03g7ghx) — The four-hour alert is field evidence for the same minimum-viable-path discipline E3 applies to agent execution: start with the smallest sufficient scope and expand only when needed.
- [How Forward Deployed Engineering is done at Factory — Eno Reyes](https://feed7.dev/p/how-forward-deployed-engineering-is-done-at-factory-eno-reyes-0zgscmd) — Kepler’s observation-led scoping complements Factory’s instrumented delivery path by identifying the right workflow pain before validators and deployment machinery are applied.

## Context Map

- Layer: agent
- Domains: coding, data
- Topics: harness-engineering, context-engineering, enterprise

## Uncertainty

- Small fixes are rarely temporary: one improvised retention script spread across a nearly **100,000-person customer** and remained in use a year later. Fast delivery therefore still needs production assumptions, ownership, and a path into the core product.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.