Strong Recommend
Sleeper
Scouting Report: Jordan Park
Staff ML Systems Engineer · Lumen Labs · San Francisco, CA · Stanford CS PhD
Scout's Take
"Plus-plus systems instincts with research credentials to match. Floor is a Staff IC at any top-5 AI lab — they're already there. Ceiling is founding engineer at the next major inference startup, or a return to academia in 5-7 years to start a hot serving-systems lab. Tenure pattern says they'll move; the question is whether for the right founder. Recruit them now while comp bands undervalue compiler-savvy serving engineers."
— VettLab, May 2026
Scouting Card
JP
Jordan Park
Staff ML Systems Engineer, Lumen Labs
San Francisco, CA
PhD Computer Science, Stanford (2022)
BS Computer Science, MIT (2018)
they / them
Est. Comp (TC)
$580 – 720k
Edge
Compiler + LLM Serving
Combine Measurables
3
Papers
1 first-author MLSys
1,847
Citations
Google Scholar
4.2k
GitHub Stars
412 in last 12mo
3
Conf. Talks
MLSys, GTC, CUDA Mtp
Projection
Floor
Staff IC at any frontier AI lab. Already in band — Anthropic / OpenAI / DeepMind would re-hire on a phone call.
Ceiling
Founding engineer / first systems hire at a $100M+ AI infra startup. Or, in 5-7 years, faculty at an R1 building a serving-systems lab.
Bust Scenario
Founder mismatch → bounces to a 4th gig in 18 months → settles for senior IC with "serial mover" reputation.
Player Comp
Technical arc resembles Tim Dettmers circa 2021 — pre-bitsandbytes-virality but post-PhD foundation. Compiler instincts more developed than Dettmers at the equivalent stage; weaker on the social-media flywheel.
Profile Dimensions
Career Timeline
2014 – 2018
BS Computer Science, MIT
TA for 6.824 (Distributed Systems) two semesters. Senior thesis on cache-coherent distributed key-value stores.
Summer 2017
SWE Intern, Google Brain
Worked on TPU compiler optimization. First exposure to large-model training systems.
Summer 2018
Research Intern, NVIDIA
CUDA kernel work for sparse linear algebra. Resulted in two cuSPARSE patches.
2018 – 2022
PhD Computer Science, Stanford (Systems Group)
Advisor: Prof. Helen Vasquez (composite). Thesis: "Compiler-Assisted Kernel Specialization for Mixed-Precision LLM Inference." Three publications during PhD; one MLSys best-paper runner-up.
Aug 2022 – Mar 2024
Member of Technical Staff, Anthropic
Inference team. Contributed to latency improvements on the Claude 3 family. 19-month tenure.
Apr 2024 – Present
Staff ML Systems Engineer, Lumen Labs
Employee #15 at a Series B AI infra startup. Building the next-gen serving stack: custom CUDA kernels, distributed inference, mixed-precision compiler.
Risk Factors
Watch-Outs Before Engaging
- Short current tenure. 13 months at Lumen. Anthropic was 19 months. Pattern reads "follows the interesting work, not the title."
- Public output gap. Twitter dormant since Feb 2026; no new blog posts since November. Could be heads-down, could be checked-out, could be incubating something.
- Side-project chatter. Mentioned a "weekend thing" on the Latent Space podcast (Q1 2026). Unverified — but if it's a stealth founding move, that affects flightability and intent.
- No people-management track record. All IC. If the role expects org-building, this is the gap to probe in interview.
- Why-leave-Anthropic question. The 19-month exit was unusual for a frontier-lab MTS in good standing. Get the real story before the offer stage.
Network Connections
Prof. Helen Vasquez
Stanford CS Systems Group
PhD advisor. Hall-of-fame systems researcher with deep ties across NVIDIA, Google, and the AI lab circuit. Strongest single warm path.
Thesis Advisor
Anthropic Inference Alumni
2022–2024 Cohort
Roughly a dozen ex-colleagues now scattered across AI labs and infra startups. Tight-knit cluster. Backchannel reference quality is high.
Former Co-workers
vLLM / Mosaic Contributor Circle
OSS Collaboration
Active in PRs and design discussions across two of the major LLM serving projects. Path in via maintainer DMs.
OSS Network
"Kernel Mafia" SF
Informal Cluster
Loose group of ex-NVIDIA, ex-DeepMind, ex-Anthropic kernel writers in SF. Meet at CUDA Engineers Meetup and private dinners. High-density warm intros.
Informal
Recruiter Notes
Likely Motivations
- Wants to ship serving systems that move the frontier, not maintain mature stacks
- Compiler-systems intersection is rare and they know it — will pick the role with the best technical canvas, not the highest comp
- Founding-engineer narrative will resonate harder than Staff-at-FAANG
- Academic re-entry is a real outside option in 5-7 years; reference it when discussing what the next 24mo enables
Compensation Expectations
- Current TC at Lumen estimated $580-720k (cash + equity at Series B prices)
- Frontier labs would pay $700k-1M+ for a Staff IC with this profile today
- Founding-engineer role: cash flexibility, but expect 1-3% equity at seed/A or it won't compete with frontier-lab base
- Sign-on bonuses move them less than equity refresh structure
Warm Intro Paths
- Prof. Vasquez (Stanford) — if you have a Stanford CS systems contact, this is the highest-conversion intro
- Ex-Anthropic inference colleague — backchannel a current Anthropic IC; the "should I take a call from X" filter is tighter at frontier labs
- vLLM core maintainer — OSS collaboration history opens that door
- CUDA Engineers Meetup organizer — SF event circuit, in-person warm path
Information Gaps
- What's the rumored side project? Could be hobby, could be stealth co-founder move
- Reasons for leaving Anthropic at 19mo — performance, team change, or proactive pull?
- Stated comp expectations not public; estimate is range-anchored, not source-confirmed
- Management appetite unknown — will they accept a tech-lead role that includes hiring?
Full Intelligence Report
Executive Summary
Jordan Park is a Staff ML Systems Engineer at Lumen Labs (Series B AI infrastructure startup), employee #15, in San Francisco. PhD from Stanford CS Systems Group (2022) under a hall-of-fame advisor; BS from MIT (2018). Spent 19 months at Anthropic on the inference team before joining Lumen in April 2024. Specialty is the compiler-systems intersection — mixed-precision LLM serving, custom CUDA kernels, distributed inference. Three peer-reviewed papers (one first-author at MLSys, two second/third-author at NeurIPS), 1,847 citations, and a 4.2k-star OSS project. Highly recruitable on signal: short current tenure, public output gap since February, podcast hint of a stealth side project. The technical core is plus-plus; the open questions are management appetite and what they'd actually move for.
Profile
| Name | Jordan Park |
| Pronouns | they / them |
| Current Role | Staff ML Systems Engineer at Lumen Labs (employee #15) |
| Location | San Francisco, CA |
| Education | PhD CS, Stanford (2022, Systems Group); BS CS, MIT (2018) |
| Specialty | LLM inference systems, CUDA kernels, mixed-precision compilation, distributed serving |
| Career arc | MIT → Stanford PhD → Anthropic (19mo) → Lumen Labs (current, 13mo) |
Technical Footprint
Selected Publications
- FlashKernel: Speculative Decoding with Compiler-Aware Tile Layouts. MLSys 2024 (1st author). 642 citations.
- Sparse Attention via Block-Hash Routing. NeurIPS 2023 (2nd author of 5). 891 citations.
- Memory-Efficient Activation Recomputation for Large MoE Models. NeurIPS 2022 (3rd author). 314 citations.
Open Source
- kernelzoo — maintained CUDA kernel library, 4.2k stars, active issues, used by at least two AI labs in production (per public blog references)
- vLLM — recurring contributor; 14 merged PRs over the past 18 months
- Mosaic Composer — 6 merged PRs around mixed-precision training
Talks
- MLSys 2024 — FlashKernel paper presentation
- GPU Tech Conference 2024 — "Compiler-Aware Tile Layouts in Production Inference"
- CUDA Engineers Meetup SF 2025 — "Mixed-Precision Inference: What Actually Wins"
Assessment: Depth over breadth. The publication record is small but every paper is substantive at venues that matter for systems work. The OSS work signals taste — they pick high-leverage projects and contribute material features, not drive-by typo fixes.
Network & Connections
Stanford Systems lab is the academic anchor; the Anthropic inference cohort is the industry anchor. Both are dense, opinionated, gossip-prone groups — backchannel references will be candid and cheap to obtain. The "kernel mafia" SF cluster (informal, 30-50 senior engineers across NVIDIA, Anthropic alumni, ex-DeepMind) is the realistic warm-path layer for outreach. The vLLM contributor list is a public-but-underused on-ramp.
Digital Presence
| LinkedIn |
Active — 2.1k connections, sparse posts, recommendations from Stanford and Anthropic alumni |
| GitHub |
Active — @jordanparker (composite handle), 4.2k cumulative stars, 412 in last 12 months |
| Google Scholar |
Active — h-index 8, three first/co-author papers, 1,847 citations |
| Personal Blog |
Quiet — jordanpark.dev. Four deep-technical posts; last post November 2025. |
| Twitter / X |
Dormant — 6.4k followers; no posts since February 2026 |
| Podcast Appearance |
Confirmed — Latent Space, January 2026 (~52 minutes, mentioned a "weekend project" without specifics) |
Reputation & Red Flags
Positive Signals
- Stanford CS Systems Group is blue-blood; advisor lineage carries weight
- MLSys first-author paper — among the highest-signal venues for serving-systems work
- Production-quality OSS used by other AI labs (rare combination of research credibility + shipping discipline)
- Backchannel reception in Anthropic inference circles is reportedly strong
Concerns / Red Flags
- Two short tenures in a row (19mo, then 13mo and counting) suggest someone who optimizes for technical interest over org loyalty — could be a feature, could be a flag
- The Twitter/blog silence since early 2026 is the strongest "something is up" signal — heads-down on Lumen work, or quietly building something
- No people-management exposure. If the next role expects them to lead 5+ engineers, that's an unproven dimension
Recruiter Notes
The pitch that lands isn't "more compensation" — it's "the technical canvas you can't get at a frontier lab." Lumen is small enough that they probably already have outsized scope; the move would have to either (a) shrink the company further toward founding-engineer scope, or (b) widen the canvas materially (e.g. a frontier lab giving them a whole serving-systems org to design from scratch).
Open with the technical problem you're solving, not with the comp number. Get to comp on call 2 once they've decided the work is interesting. If you're a recruiter representing a founder, lead with the founder — this person will pattern-match on whether the founder thinks like Vasquez or like a McKinsey graduate.
Reference checks: Anthropic inference alumni will give the truth fast. Stanford lab alumni will give the truth slowly. Don't ask Lumen colleagues until very late in the process — the company is small enough that the question signals intent.
Format Notes