FM-os
public🛠️ FM-os: the living, SLM-first map of foundation-model operations — pre-training, post-training, fine-tuning & RL. Curated repos, courses, papers & jobs, auto-refreshed weekly.
AI Founder · Inventor · Investigator · Investor · MMA and Outdoor Recreator
I build self-improving AI systems in the open, and each one has to prove itself before I call it done.
33 projects active in 90 days · last push 7 days ago3 of 6 résumé claims corroborated · 2 unverified216 long-form pieces
🤖 Ask my agent — who are you?
Making My Ceilings Your Foundations for a 12XFuture Together
I build self-improving agentic systems — operating systems for people and companies — and I build them in the open. By day I lead strategic collaboration across frontier AI ecosystems and ship agentic AI into enterprise production as an Advanced AI Scientist & Engineer at Accenture; nights and weekends I compound a fleet of agent-native “OS” products and teach others to do the same.
33 projects active in 90 days · last push 7 days ago3 of 6 résumé claims corroborated · 2 unverified216 long-form pieces
Compounding everything
Begin from the End — Result-Oriented System Engineering (ROSE): every venture starts as a one-page end-state — the 12X result, its metric, its date — written before anything is built. Then the arithmetic takes over: (1+0.12)^22 ≈ 12.1, so a 12X future is ~22 compounding loops, each moving a named metric by ≥12% now, measured — or it wasn't a loop, it was motion.
T Transferable & Transformative — End-state-first plus the compounding identity (1.12^22 ≈ 12.1) transfers to revenue, research, and relationships alike — any 12X dream decomposes into a designed finish line and ~22 measured loops.
R Reusable & Refinable — Two reusable artifacts: the one-page end-state (result, metric, date, owner) and the loop contract (name the metric, move it ≥12%, show the measurement) — both refinable every re-plan.
U Understandable & U-loop — One breath: begin from the end at 12X, then demand 12% from every loop — because anything less never compounds there.
E Experienceable & Experimentable — Write the end-state page for your current project today, then run one loop this week and read its measured delta. No +12% reading = motion, not progress.
For you — You never start without the one-page finish line — and you hold both numbers at once: 12X on the wall, +12% measured in every loop.
For an agent — A skill that rewrites any goal as its one-page end-state (metric, date, owner), plus a workflow gate: a loop only counts when its named metric shows the +12% delta — measured, never claimed.
Big dreams are earned three ways. Amazing art: elegant beauty that strikes awe — simple form, profound truth. Great science: a falsifiable hypothesis. Deep engineering: a tested, shipped system. Every bet carries all three, or it's a wish — then vision sets the heading, telemetry compounds it, and each week one open gap gets closed.
T Transferable & Transformative — Art-plus-hypothesis-plus-system transfers across domains: the same triple that ships a trading engine ships a biomarker pipeline — and awe travels farther than any spec sheet.
R Reusable & Refinable — Three reusable artifacts: the falsifiable claim, the test suite, and the weekly gap-review — every failed hypothesis still compounds the map, every closed gap raises the floor.
U Understandable & U-loop — One sentence: earn the bet three ways — beauty so elegant it strikes awe, truth so simple it runs profound, a system so tested it ships — then let the destination pull while the sensors correct.
E Experienceable & Experimentable — Write your biggest bet's falsifiable claim and its failing test today; close one open gap this week and check that next week starts from higher ground.
For you — You hold every big dream to three bars — elegant beauty that strikes awe, a claim that could fail, a system that ships — then let the future set the heading and close one gap a week.
For an agent — A plugin that refuses to open a build until the bet carries both artifacts (the hypothesis that could fail, the test that would show it) — feeding the telemetry-fed re-solver workflow that recomputes the best next moves from actuals.
The base case and the multiplier: faith and family are floors, not variables — and friends don't just get encouragement, they get enablement: shipped tools, real introductions, honest coaching.
T Transferable & Transformative — The ordering — God, family, friends — transfers to every decision at every scale; it is the one constraint set that never gets re-optimized away.
R Reusable & Refinable — Returned to daily, not outgrown: the base case is re-affirmed each morning and each high-stakes decision re-checks against it.
U Understandable & U-loop — Recursion needs a base case: this is the value that stops the optimization from eating the optimizer.
E Experienceable & Experimentable — Experienceable in the calendar and the ledger: time with God and family is scheduled first, and each month one friend receives enablement you can point at — a shipped tool, a real introduction, honest coaching.
For you — You keep faith and family as floors, not variables — and you enable friends with artifacts, not just applause.
For an agent — A guardrail hook: before any high-stakes autonomous action, surface who this serves — God, family, friends — and stop when the answer is nobody.
Built in the open (last 12 months)
🛠️ FM-os: the living, SLM-first map of foundation-model operations — pre-training, post-training, fine-tuning & RL. Curated repos, courses, papers & jobs, auto-refreshed weekly.
The god's-eye view of what top-rated repos actually teach — verified, weekly-synced, machine-audited. Three provenance-pinned pillars: transferable knowledge, agentic tooling, community.
🤖 The most comprehensive, community-driven resourcefor Automated AI Research — papers, tools, people, lbs, roadmaps. Featuring Recursive Lab, Sakana AI, DepMind & beyond.
🕸️ the Graph Engineering Operating System — the model finds text; the graph finds reality
🧪 the Evaluation Operating System — OEC (Observe → Evaluate → Control) for foundation models, agents, and business objectives
Any design intent in (text, picture) → execution-ready 3D blueprint out — construction- and 3D-print-verified. Game design · simulation · architecture (3DCP+AI) · interior · landscape.
ARCHIVE (private, do not publish) — pre-rewrite history of eval-anything. Superseded by the public repo; retained only because it contains an unreleased sibling product's walkthrough in old commits an
🧬♻️ An AI-native, build-in-public compounding loop for aging science: falsifiable question → open data → public verifier → honest write-up → share → compound. No wet lab required.
Build, grade, improve & accountably ship any agent app from one sentence — a tested OOP engine with a closed RefineLoop and swappable seams. (Private.)
Agentic personal portfolio — Next.js + CopilotKit site whose on-page AI agent answers about my work, powered by a free-LLM survival chain (NVIDIA NIM→Groq→Gemini). Built in the open.
Private: decision engine + agentic store + founders' ops for a Bay Area outdoor-gear startup
Play-Engineering: one loop (want · play · toy · name · second case · crit · teach) to master any subject, age 7 to grad — evidence-gated, run-measured. Private.
SAW strategy: the plan as an Observe-Evaluate-Control loop (private)
Forward-Deployed-Engineer OS — an operating system for shipping agentic solutions into the field.
PRIVATE — Collaboration Intelligence Layer: pre-registered, closed, compounding loop per invited frontier company
Stanford Eden core team: pray, brainstorm, plan, act, observe, learn — one loop with a steward agent
PRIVATE — staffing instrument for the Accenture x NVIDIA/Kumo relational-ML team: every seat is a pre-registered experiment about what predicts delivery
Private office for an AI-native founding team: Nextcloud + OnlyOffice as code, AI agents as scoped teammates (MCP).
18 shown · swipe or use ◀ ▶
Long-form on LinkedIn
Two sermons by Pastor Robert Burbatt on the Garden of God, read through play-engineering, Shepherd's Academy of War and Destination, Not Toolbox. Eden was a meeting place before it was a park, and the man put there had two jobs: cultivate it and keep it. A landscape is kept for the viewers, an orchard for the yield, a garden for presence. The first attack in the Bible reads like a prompt injection against a spec that had already drifted. With a free 3-minute Garden Check.
Agentic software is not a SaaS app with a chatbot: you set the destination (goal, hard limits, what good and bad look like), the machine rules out every option that breaks a limit, ranks the rest and shows a few different ones, and you choose. Built on a real refactor of a founding team's office, with a playground to break and the principle under it: max-calling, min-cost.
Observe, evaluate, control, at every scale at once: a unit test, a quality floor that only rises, an outside review, a weekly loop and dated tripwires. Built on a codebase that graded itself from 65 to 97, then needed a live run and a reviewer to catch eight bugs its tests missed (fixed: 98), with a dimension ladder from a single number to a fractal.
Truth, love and beauty behave like three faces of one character: a weak one is revived by strengthening the other two, as long as they stay coupled and something keeps feeding them. A toy model you can run (pushing love alone left truth at 0.04; coupled, it came back to 0.59), three receipts from moving a live team office, and the capstone: all three principles on one system.
A six-week first AI-agent workflow for four industries where mistakes are most expensive: financial services (SR 26-2 replaced SR 11-7 in April 2026), life sciences, industrial manufacturing, and software and telecom. Each has a human boundary and a paired speed-and-quality measure. In three of the four, the trial stays out of production.
Six practices for pitching AI agents in 2025–26, built around one idea: pick one stuck workflow and let a paired measure agreed in advance do the selling. The four Rs of stuck work, the person who can say no, and why the best trials of AI coding tools disagree (experienced developers were 19% slower and felt 20% faster).
Seven decades-old principles, from SPIN selling to preregistration, turned into four moves that fit in a minute: listen to the customer's words, locate the stuck workflow and its cost, limit a trial they can watch, and prove it with evidence agreed in advance. Plus why slide 7 is where buyers stop listening.
A complete world model small enough to read in one sitting is six lines in CS229 section 19.3: count what the world does, plan with value iteration on what you counted, act, repeat. Play it as a shepherd learning a wolf (learning lifts it from 17–37% of wolves turned away to 87–100%, seed 1), then climb to Kalman, Dyna, World Models, Dreamer, MuZero and Genie, with every famous claim's hedge kept.
Discrete math counts with gaps; continuous math measures without them. Four turns where each borrows the other (limits, analytic proofs of discrete facts, sampling and computation, relaxation and rounding), the condition that triggers each, the technique, and the motive: truth, love or beauty. With three live demos and runnable code.
For 20-year-olds who code: the child's idea that a number is not its name is the first abstraction in software. Hume's Principle, log-length positional codes, zero-free base 26 in your spreadsheet, two's complement, floating point, Unicode digits, and an invariance test you can add to your own code.
For 15-year-olds: a two-hand base-5 counter, a one-division proof that every number has exactly one spelling in any base, a zero-free base ten where one million is 99999κ, the test that separates a number from its notation, and a body map of a number system that says where it breaks.
For 11-year-olds: invent your own number system, race it against tally marks, 正, Roman and Egyptian numerals in a live playground, and see why each new way of writing fixed the last one's problem, why zero had to hold a place, and which facts are about numbers and which about how we write them.
Proof, not claims — a résumé audited against real artifacts, then closed into a verified one
Verify it yourself
Skeptical? Paste a résumé — mine, or any — and watch each claim audited live against real public GitHub. Honest by design: unprovable claims come back unverified, never rubber-stamped.
Your run is shown here in your browser; it doesn’t change the published proof.
3/6 claims corroborated by real artifacts, 1 partial, 2 need an external source.
top gap: Teaches a Silicon Valley AI-architect cohort. — A cohort syllabus crediting Paul as instructor, or a participant testimonial.
Sample receipts (built from public repos). Ask the agent to verify a real résumé. · source: “(Sample receipts — paste your own to the agent: “verify this résumé: …”.) Paul Jialiang Wu…”
What people who've worked with me say — imported from LinkedIn, or written here (I approve each one; you can post yours to LinkedIn in a click)
“I worked with Paul at Genentech where he was the Principal Data Scientist on my team. His technical knowledge was always astounding and his understanding and direction for the possibilities of Machine Learning using our available data gave our team direction and the confidence to explore new projects. He was a pleasure to work with and I recommend him for any team in need of a technical data/ML lead.”
Senior Machine Learning Operations Engineer @ Genentech - People Insights Roche
Rajesh worked with Paul Jialiang on the same team
“I have been working with Paul on a GraphML project in the past 6 months. Paul is a passionate leader with a solid technical background. He always try his best to inspire and influence others, and to encourage others to share their ideas. Besides, he can build and organize the workflow properly. There is no doubt that Paul will be a great asset to your team.”
Data Scientist - People Insights at Roche
Paul Jialiang was senior to Zhao but didn’t manage Zhao directly
“Paul has a great energy and is a good story teller. He leaded a focus group with high-tier people and we worked together to put our business to the next level, 10X-style.”
Fondateur de CAIRN-CREA | Je fais du contenu organique & pub qui convertit pour marques DTC
Louis worked with Paul Jialiang on the same team
“I am very glad to have spoken with Paul — he's a champion for self-improvement and mental health, and is also a great person to talk to!”
“I had the pleasure of attending Paul's presentation on GraphML in a conference, he is an amazing Data Scientist and a brilliant instructor of latest technologies in Data Science and Machine Learning. He explains very complex concepts, turns challenging topics look easy by breaking them into clear and concise steps. He tries to teach tough concepts by correlating them with real world scenarios which makes him a great story teller. His principle of improving 10x in 3 months is astonishing. He can guide people with any years of experience to be successful in their Data Science journey. I hope to work with him very closely and continue learning from him in the future. Paul would be a great asset for any organization!”
“I had the pleasure of working with Paul at Galvanize, he is a brilliant and gifted Data Scientist who truly understands how to get the best out of people. Paul always made sure to make the students feel at ease, even when he was teaching challenging topics. He always made sure that his students got most of the class and he was always fun to co-teach with. The biggest strength in Paul, we all in the team admired was his ability to deal with conflicting priorities in high-pressure situations. He never lost his cool and was always a people's person. Paul would be a true asset for any company who need a game changing Data Scientist.”
“Paul is one of the best instructors in data science. He is super knowledgable and breaks down complex concepts into clear and understandable topics. He is a dedicated teacher and I learned so much from him in a short amount of time and I hope to continue learning from him in the future.”
Senior Data Scientist @ Ent Credit Union | AI/ML
Paul Jialiang was senior to Bahar but didn’t manage Bahar directly
“Paul was a substitute instructor for our Data Science Immersive program and he was absolutely amazing! He was easy to work with, the students loved his teaching style, and he showed a great depth of Data Science knowledge and teaching ability. I hope to collaborate with him again in the future and I would highly recommend him for any position on your team.”
Strategy & Operations | Tech Partner Ecosystem Development
Kristen worked with Paul Jialiang but on different teams
“It was my pleasure to be in a class that his teacher is Paul Jialiang, he has all the skills that I feel it is really important that the teacher should have, it was easy for him to make the concept clear and what I really like about him, is he make us feel good about ourself spotting the light on our hard work and what we know and learned,, I felt happy after your class thanks again Paul”
9 shown · swipe or use ◀ ▶
Score any job against past experience · current skillset · future mission/values/vision — held to a golden-set accuracy
Score a role against me
Paste a job-posting URL (Ashby/Greenhouse/Lever, fetched live) or the JD text. The agent scores fit across past experience, current skillset, and future mission/values/vision — and tells you, honestly, where it doesn’t fit.
Why trust this
8/8 within one band · 100%This scorer is itself held to the standard it holds JDs to: it was run over a 8-example golden set with human-assigned fit labels and agreed within one band 100% of the time (6 exact). Verify, don’t vibe.
No role scored yet — paste a posting above to see the fit breakdown.
Knowledge graphs + agentic tooling (skills, plugins, workflows, hooks, bundles) I've built — full for me, a summary card for visitors
AnyAgent is a tested, object-oriented engine designed to build, grade, improve, and ship agent applications from natural language.
Owner-only knowledge graph + tooling. Unlock owner mode to view the full package.
Retrieved, not guessed: 3 competency clusters, 9 concept/tool nodes grounded in real sources (9 Wikipedia/GitHub definitions), and 6 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.
Retrieved, not guessed: 6 competency clusters, 23 concept/tool nodes grounded in real sources (21 Wikipedia/GitHub definitions), and 8 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.
Retrieved, not guessed: 5 competency clusters, 17 concept/tool nodes grounded in real sources (15 Wikipedia/GitHub definitions), and 9 skills from the ESCO taxonomy — off-domain hits filtered out. Every node carries a real definition; edges are real relationships where an open KG had them.
Transformers scale COMPUTE (Mixture-of-Experts routes tokens to experts) but have no native primitive for looking knowledge UP. DeepSeek's Engram adds that missing primitive: it modernizes classic N-gram embeddings into a deterministically-addressed table with O(1) lookup — a second, complementary axis of sparsity (static memory) alongside MoE's conditional compute. The paper finds a U-shaped scaling law: for a fixed budget there's an interior optimum splitting capacity between compute and memory — going all-in on either side loses. Under iso-parameter and iso-FLOPs constraints, Engram-27B consistently beats MoE baselines on knowledge, reasoning, code, and math. Mechanistically, offloading static recall to Engram frees the early layers from pattern reconstruction, preserving effective depth for reasoning — and the huge embedding tables offload to host memory with minimal inference overhead. Why it matters for you: 'conditional memory as a sparsity axis' is a transferable design move — separate what a system should COMPUTE from what it should LOOK UP, and size each.
Four growth vectors — deepen · widen · lengthen · heighten — plus who to reach. Drafted for approval.
Four growth vectors
↓ Deepen
more fundamental, seminal — to the roots
first-principles / foundational research
↔ Widen
new applications, features, markets
Ansoff · Innovation Ambition Matrix
→ Lengthen
evolve it to robustness, scale, commodity
McKinsey Three Horizons · Wardley evolution
↑ Heighten
generalize, abstract, compress the mechanism
abstraction laddering · compression (MDL)
Further develop and refine the existing self-improving agentic operating systems to increase their efficiency and capabilities
first step: Integrate idle detection and SMARC output-quality verification from sos into loop-engineering-anything
The collaborator candidates were selected based on their relevance to the widenInterests and verified strengths of the builder.
Drafted for your approval — nothing is sent automatically. · model groq:llama-3.3-70b-versatile
A fifth vector — emergent projects from bisociation across the fleet. Scored in code; drafted for approval.
secure-anything-core × agentgate × company-os
An operating system for AI-native companies where every agent action, decision, and code change is routed through a secure, auditable gate and enforced by a lean, policy-driven security core, providing unparalleled transparency, compliance, and control over an autonomous workforce.
pivot: Policy-driven trust and accountability for AI agent fleets in a corporate context.
first step: Implement `agentgate` as the primary router for all agent actions within `company-os`, with `secure-anything-core` defining the initial set of access control policies.
anyagent × loop-engineering-anything × eval-anything
A comprehensive operating system that takes a one-sentence agent app description, turns it into a self-improving CLI via a generate-judge-refactor-rejudge loop, and rigorously evaluates its performance against objectives using an OEC framework, creating an autonomous agent development pipeline that continuously optimizes agent behavior and code quality.
pivot: Closed-loop, OEC-driven self-improvement for agentic applications.
first step: Integrate `eval-anything`'s OEC metrics as the 'judge' component within `loop-engineering-anything` for an agent built using `anyagent`.
GodView-anything × rsi-os × loop-engineering-anything
A meta-research agent that, using verified knowledge, identifies gaps and trends, then leverages tools for automated research, and finally continuously refines its research strategies and output quality, becoming a truly autonomous scientific discovery engine.
pivot: Automated, self-improving knowledge discovery and hypothesis generation.
first step: Develop a `loop-engineering-anything` agent that uses `GodView-anything` to identify a specific research gap in `rsi-os`'s curated papers and proposes a new experiment.
FM-os × eval-anything
An autonomous system that not only maps FM operations but also actively monitors their performance in real-time, identifies deviations from objectives, and triggers corrective actions or fine-tuning pipelines, creating a truly adaptive and self-improving FM lifecycle.
pivot: OEC (Observe → Evaluate → Control) for Foundation Models.
first step: Integrate `eval-anything`'s OEC loop with `FM-os`'s curated list of fine-tuning methods, allowing automatic triggering of fine-tuning based on evaluation metrics.
agentic-marriage × marriage-counseling-coach × havenandheaven
A private, AI-native sanctuary that provides a verified route from marriage pain points to committed steps, complemented by a personalized AI coach for healthier marriages. This creates a holistic platform offering practical guidance, emotional support, and a safe space for growth and healing in relationships.
pivot: Empathetic, verified, and holistic AI support for personal relationships and well-being.
first step: Develop an agent in `marriage-counseling-coach` that can ingest a pain point identified by `agentic-marriage` and suggest a relevant, empathetic conversation starter or resource from `havenandheaven`.
money-os × career-os × achiever-os
A zero-trust, local-only AI operating system that holistically manages an individual's life. It combines `money-os` for financial consciousness, `career-os` for autonomous job search and skill upgrading, and `achiever-os` for goal setting, measurement, and gap closure, creating a compounding, AI-native engine for personal wealth, career advancement, and overall achievement.
pivot: Integrated, AI-driven personal growth across career, finance, and achievement.
first step: Create a dashboard in `achiever-os` that pulls career progression metrics from `career-os` and financial milestones from `money-os` to provide a unified view of personal goal attainment.
design-anything × physical-ai-native-os
A system that takes high-level design intent (text/picture) and autonomously generates execution-ready 3D blueprints, then orchestrates the entire physical fabrication, assembly, and deployment process through AI-native operations, including teleoperation and incident triage, enabling rapid, on-demand physical product realization.
pivot: AI-native design-to-physical-deployment pipeline.
first step: Develop a module in `physical-ai-native-os` that can interpret a `design-anything` generated 3D blueprint for a simple object (e.g., a chair) and generate a bill of materials and a basic assembly plan.
longevity-loop × lifescience-os
A compounding loop for aging science that is fully operationalized by an agentic AI for Life Sciences. This system would autonomously generate falsifiable questions, source open data, run public verifiers, pressure-test verdicts, and deliver cited briefs, culminating in a feedback loop that continuously refines research and translates findings into actionable, personalized health journeys.
pivot: AI-native, falsifiable scientific discovery and application in life sciences.
first step: Develop an agent within `lifescience-os` that can take a falsifiable question from `longevity-loop` (e.g., 'Does X compound affect Y biomarker?') and identify relevant open datasets for analysis.
Emergent projects from bisociation across your fleet (Swanson A–B–C + conceptual blending), scored in code. Most combinations are noise — the ranking is the value. Drafted for you; nothing auto-built.
One baseline, ten tailored, zero invented sentences
Every tailored résumé is a projection of one baseline: it may select, reorder, compress and reframe — never introduce a claim. What varies between them is sorted by how fast it changes.