XNews

Forward Deployed Engineer (FDE) Best Practices Analysis

Forward Deployed Engineer (FDE) Best Practices Analysis

A detailed analysis by Kepler CEO Vinoo Ganesh and an Anthropic engineer defines FDE as an engineer who returns platform-level insights from customer deployments. Key practices include reporting to product not sales, building a platform that accumulates verified domain knowledge, and avoiding consulting traps. A case study shows a Parquet conversion stalled by a missing viewer tool.

Key insights

FDE definition and pollution of the term

Kepler CEO Vinoo Ganesh noted that the term FDE is heavily polluted: at an a16z FDE Fellowship dinner, FDEs from Snowflake, Anthropic, and startups described three different jobs—pre-sales engineer, quota-carrying sales who can script, or on-site consultant filling product gaps. Ganesh gives a strict definition: FDE is an extension of the product team; solving customer problems is the means, but the true purpose is returning insights about what should be built into the platform.

The 14TB memory incident illustrating the cost of missing frontline knowledge

In 2013, a Palantir Phoenix project crashed because an empty timestamp fell to epoch (1970-01-01), causing stale logic to request about 2.3 million keyspaces, resulting in server OOM that required 14TB of memory to restart. The root cause was not a code bug but that nobody had stood in the customer's building running the system against real production data. This incident visceralized the FDE mindset.

Methodology: collecting 'nouns' and 'verbs'

FDEs should collect 'nouns' (core objects the enterprise treats as real, e.g., a position, trade, or counterparty) and 'verbs' (how nouns flow—how trades are booked, what conditions must be met before closing, who signs off on exceptions at 11 PM). This knowledge is 'lived', not written, existing in six employees' heads and a four-year-old spreadsheet. It is valuable precisely because it cannot be elicited by interview.

Parquet case: hidden true reason for resistance

A startup spent nearly a year failing to convert a customer from CSV to Parquet, blocked repeatedly by a data quality engineer whose reasons kept changing. An FDE observed her on-site and discovered the real reason: she visually inspected CSV files on a Windows laptop, and Parquet had no viewer—migration would remove her only quality inspection tool. The team built a Parquet viewer overnight, and she approved migration two days later, reducing pipeline runtime from 17 hours to 2 hours. This true motive could not have been discovered in interviews.

Output must be product, not customer satisfaction

A solution architect succeeds when the customer is happy; an FDE has a higher bar. If at project end there is only a happy customer and no upstream change to the platform, the FDE has failed. A cautionary example: an afternoon's script called vinoo.groovy was still running a year later on a customer with 100,000 users. The lesson: 'every shortcut you deliver becomes your debt.' Discipline means deliberately deciding which fixes should go into the platform and which should be actively deprecated.

Kepler's four practices

1) FDE reports to product line, not sales—this structural choice determines whether the focus is closing the current deal or enabling the next deployment. 2) Let invariants define the platform: for financial customers, numbers must be provably correct, so provenance is the platform skeleton. 3) Deployments reveal where the platform is 'too narrow', not what features customers want—three companies each bypassing the same limitation is a real signal. 4) Product leverage buys the right to fail cheaply, allowing 'four mistakes a month' rather than one expensive guess per quarter per customer.

Moat: accumulated, current, verified understanding of vertical enterprise operations

Ganesh argues the moat is not the model (rented, getting cheaper), talent (price already discovered), or knowledge of a single customer (extraction almost free). Instead, it is accumulated, current, and verified understanding of how enterprises in a given vertical actually operate, packaged in a platform that keeps it up-to-date and provable. The first deployment is anecdote, the tenth is a pattern. Business drift makes outdated models silently fail. Plausible and truly correct coding look identical until something breaks.

FDE 101: when to use FDE and what makes a good FDE

Anthropic engineer Kevin Bai's FDE 101 noted that Palantir achieves $4M average contract value versus $1.2M for ServiceNow, $600K for Workday, thanks to the FDE model. FDE is needed when selling a technically complex product to non-technical buyers. The prerequisite is a reusable platform—without it, engineers write custom code from scratch, becoming unsustainable consulting. As AI makes software highly customizable, more companies may need FDEs. A good FDE is a software engineer you trust to directly face customers.

Source posts