AI-assisted engineering
Agent workflows with real guardrails: specs before code, independent review, bounded commands, cost tracking, and a learning loop so the same mistake isn't made twice.
z-nile · 8.1% rework
Staff software engineer · Developer productivity × AI
Philip Bergen
I make engineering organizations faster. For 16 years at Salesforce and Meta that meant tools thousands of engineers chose to use. Now it also means putting AI agents in the loop, and keeping what they ship tested, reviewed and measured.
Open to staff and principal roles · Sandpoint, Idaho
- Meta90%less Reliability Time Wasted in the Ads ML training pipeline
- Meta60%+fewer operations in the ML life cycle, saving 20,000 hours each half
- Salesforce750→15,000engineers on the developer platform I owned
- Salesforce97%of developers used my tools, by choice
- z-nile8.1%rework rate across 136 agent-delivered issues
What I do
Human time is the expensive resource
I started by making code fast, then systems, then teams. The work is the same at each layer: find where time is wasted, measure it, and remove it.
Developer productivity platforms
Setup, sync, build, CI/CD and version control for very large monorepos. I measure adoption and time saved, and I build tools people pick without being told to.
Salesforce · 97% adoptionReliability and performance
From a hand-built C++ database that queried tens of billions of rows in seconds on 2004 hardware, to the reliability program for Meta's Ads ML training pipeline.
Meta · 90% less time wasted
Experience
Thirty years, one direction
-
Now
Advisor and contract engineer · Doubling
Advising and occasional contract work, alongside building z-nile.
-
2023 – 2026
Senior Staff Software Engineer (IC7) · Meta
ML engineering productivity for the Meta Ads algorithm: automation, tooling and workflows. I drove the reliability program and reduced Reliability Time Wasted by 90%, shortening the ML delivery cycle by two months. That translates to 4–6% more revenue on a $20B-a-year ads algorithm. Tech lead for the ML model metrics service and tooling used by about 5,000 engineers, which cut the number of operations by over 60% and saved 20,000 hours per half.
Python · MySQL · Hack · Sapling · Buck · Rust · Go · Claude -
2010 – 2023
Principal Architect (VP) · Salesforce
Owned developer productivity as engineering grew from 750 to 15,000: tooling, virtual workstations, version control and CI/CD. 97% of developers used my tools voluntarily. Setup went from weeks to an hour, sync from hours to minutes, and build success from 40% to 95%. Helped migrate from Ant to Maven to Bazel. The savings are in the thousands of engineer-years.
I designed and wrote o4, a highly parallel Perforce front end that made sync 100× faster while guaranteeing a perfect copy of the 500 GB codebase for 7,500 monthly users. It is open source: github.com/salesforce/o4.
Python · Bash · Perforce · Git · Bazel · Maven · Jenkins · Docker · OpenStack · AWS · GCP · Java · C/C++ -
2000 – 2007
Founder, R&D · Instadia
Web analytics SaaS, big data before the term. My hand-built C++ database handled billions of rows a month and answered online queries over tens of billions of rows in seconds, several thousand times faster than competitors, which is why the business ran profitably. Acquired by Omniture in 2007.
C++ · Java · Python · MySQL · PostgreSQL · Linux -
1998
Systems Engineer · Rovsing
Left university to build control systems for the Ariane 5 rocket boosters. Also made the flagship condition monitoring system 10× faster for its deployment at the Chernobyl nuclear power plant.
SHARC assembler · C++ · VxWorks · Matlab
Earlier roles and education
- 2009 – 2010
- Owner, Philip Bergen Consulting. iPhone development, mostly for a well-known Danish pharmaceutical company.
- 2007 – 2009
- CTO, Wiinz. Backend and control UI for a high-speed sports betting arbitrage platform.
- 2007
- Chief Software Architect, Premitech. Getting sales, product, support and development to trust each other and work together.
- 2007
- Omniture, after the acquisition. Ported the Instadia system and saw my staff through the transition.
- 1999
- Systems Engineer, Adaptive Systems. A bank teller system and its printer drivers.
- 1999
- Systems Engineer, Brüel & Kjær CMS. CPU board hardware design and Windows NT drivers for safety and condition monitoring.
- 1997 – 1998
- Teacher and lab engineer, Blekinge Institute of Technology. Taught low-level programming and microprocessor architecture. Built signal-processing prototypes; the noise cancelling ended up in Ericsson handsets.
- 1993 – 1997
- Founder and mechanic, Cykelbergen. Probably the world's first online bicycle shop. It paid for college.
- Education
- Blekinge Institute of Technology, Sweden. University Diploma in Electrical Engineering. Thesis: safety and condition monitoring of the Chernobyl nuclear power plant.
Current work · open source
z-nile takes a one-line request to a merged, tested pull request
I built z-nile to find out how far AI agents can carry real engineering work when the process around them is strict. Work moves through five stages. Each stage is a skill an agent follows, and every skill drives one engine, ./zn, which keeps the whole tracker as plain files in git.
- 1
Capture
/z-ticketA request becomes an issue. Its scope is frozen here.
- 2
Design
/z-designAcceptance specs are written before any code, then locked.
- 3
Raft
/z-raftRelated issues are grouped into one branch and one PR.
- 4
Ready
/z-readyEverything an executor needs is frozen into one prompt.
- 5
Run
/z-runAn agent builds it, a stronger model reviews it, the gate merges it.
- 136 issues and 71 rafts delivered
- 273 lessons harvested from reviews
- 24 skills
- ~20k lines of Luau in the engine
- 2,200+ commits since Sep 28
- 7 projects in one repository
- Tests come before code, and the coder can't edit them.Specs are written at design time. Review checks that the PR changed none of them.
- One clean review, then the human.A fresh agent with no shared context, on a stronger model, reviews once. Anything still open goes to a person.
- Every finding becomes a lesson.Each has a check question. Agents search them before work a lesson might cover.
- The tracker is plain files in git.No database or queue. Status, claims and spend are commits, so every agent sees the same state.
- Cost and rework are measured.Every commit records minutes, lines, tokens and model. Defects count as rework, with a target under 20%.
- Model per job, any CLI.Cheaper models capture and group, stronger ones design, build and review, across Claude, Grok and Cursor.
What colleagues say
In their words
Before working with Phil, I thought that a “10x engineer” is some kind of unicorn you'll never meet in real life. After spending three years next to Phil, on the same team, I don't know what his multiplier is but it is way over 10x.
Phil is an amazing thinker and implementor, with a particular passion for unsexy but impactful problems. He conceived of and created huge pieces of the developer productivity and workflow toolset at Salesforce, in many cases by taking a completely different approach to a problem that seemed insurmountable…
Phil has an amazing eye for driving simplicity and clarity out of complex and hard problems. He is a master of breaking down problems, and designing and delivering solutions and products incrementally and at a rapid pace.
… Phil consistently tracks, and demonstrates it all with indisputable yet easy to understand metrics. More often than not, also using the data in identifying bottlenecks and resolving all sorts of systematic and hard to identify problems.
As chief architect and director of development, Philip was recognized for his passion, strong innovative skills and for being a person not afraid of challenging prevailing thinking if he believed there was a better approach out there.
Patents and open source
Machine learning for developer tools, since before it was fashionable
Detecting an error in software
A classifier trained on pairs of buggy and corrected code that reviews new code and flags the sections likely to contain an error.
Direct build assistance
A tool that installs and configures a development environment and has built-in help for resolving build problems when they occur.
salesforce/o4
The parallel Perforce front end I designed and wrote. 100× faster sync of a 500 GB codebase, with a guaranteed-correct result.
How I work
Strong convictions, loosely held
- I solve hard problems.With tenacity and good humor, and with the analytics and documentation that keep leadership informed about where things stand.
- I use proven tools in novel ways.Approaching a problem differently is how you find the outcome that earlier attempts missed.
- I write fast code.I learned on machines so weak that nothing ran unless it was optimized. I still enjoy squeezing out the last bit.
- I care how code reads.Simple, clear code and design run faster, break less and are easier to work on together.
- I build teams where people stay for years.Trust through honesty, respect through skill and hard work, and real responsibility handed to others.
- I change my mind when the data says I'm wrong.Hypotheses get tested against the latest numbers, every time.
- Always trust
- Always write code
- Always learn
- Always open and honest
- Always inspire
- Always innovate
- Always deliver
- Always artful beauty
- Always performant
- Always responsible
- Always proactive
- Always communicate
- Always document
- Never blame
- Never micromanage
- Never assume
- Never judge
- Never ever personal
Off the clock
Powder, tractors and three boys
I live in north Idaho with my wife Emma and our three boys, who are busy exploring the limits of our patience. Winters are for skiing as a family; I hold a ski instructor certification from the Swedish Friluftsfrämjandet. In the fall I'm in the forests of Clark Fork with my boys.
In my spare time I'm building a farm, which has made me fairly good at design, carpentry, plumbing, welding and running an excavator.
- US and Danish citizen
- English · Danish · Swedish
- Conversational French
- Certified ski instructor
Let's make your engineers faster.
I'm looking for a staff or principal role where developer productivity meets AI. Email is the quickest way to reach me.
- Emailphilip.bergen@me.com
- Phone415-200-7340
- LinkedInlinkedin.com/in/pbergen
- GitHubgithub.com/philipbergen