About
Activity
209 followers
-
Kiana H. shared thisPretty cool stuff we've been working on :) Check out the post and also the interactive diagrams!Kiana H. shared thisOur co-inventor agents optimized the full AI inference stack, verifying every change with rigorous correctness checks. They worked autonomously on vLLM serving two of the largest open-source LLMs, DeepSeek v4 Pro and Z.ai GLM 5.2, on NVIDIA B200s. The agents improved throughput and interactivity by up to 16% across concurrency levels, running for about a day per level. Production inference performance depends on millions of lines of code, hundreds of configuration parameters, and runtime dynamics that shift with every workload. A fast kernel in isolation doesn’t translate into a faster system overall. Our agents optimize everything together: kernels, schedulers, load balancers, and configuration. Speed only counts if model behavior is preserved. Customizations can quietly alter outputs while still acing standard benchmarks. We check every change with statistical distribution-matching tests on token-level probabilities, across a diverse multilingual corpus. The agents also learn as they go. A lesson from optimizing DeepSeek v4 Pro on how to avoid a class of distributed deadlocks transferred directly to GLM 5.2, saving 44 minutes of wasted wait time. Each run makes the next one better and faster. Full write-up at https://lnkd.in/gnVupZ35.
-
Kiana H. reposted thisKiana H. reposted thisBuilding software for space is one of the hardest engineering problems there is. As AI becomes more capable of writing and reasoning about complex code, we wanted to explore the key question: can AI help us invent at the frontier? Here's what we found ��
-
Kiana H. shared thisReally excited to share this work from the amazing team at Asari AI! I've thoroughly enjoyed getting to learn more about designing AI infrastructure to work on complex software systems while putting this together. Check out the post! https://lnkd.in/gsVmKaqHKiana H. shared thisMigrating critical infrastructure from C to memory-safe Rust at scale is hard. We challenged our AI agents on four production libraries consisting of 40,000+ lines of code that power essential systems worldwide: • libmcs: high-precision math library that meets strict European Space Agency (ESA) standards • gzip: one of the most widely-used command-line tools for file compression • libyaml: the foundational parsing library used for YAML configuration files • zlib: used by virtually every operating system and application for data compression ➡️ Result: All four successfully translated to Safe Rust with comprehensive test validation. Our agents maintained correctness across thousands of interdependent decisions, demonstrating they can handle the rigorous system design and verification required for production engineering. We've open-sourced all four Rust translations. Development combined human-AI co-design for specifications with fully autonomous implementation by our agents. Read the full blog post here: https://lnkd.in/gWyMhxuH Open-source repos: libmcs: https://lnkd.in/gQuR25sP gzip: https://lnkd.in/ge7ArKiq libyaml: https://lnkd.in/gDhBEc72 zlib: https://lnkd.in/gygpvw4k All feedback welcome!
-
Kiana H. shared thisBeen working on this for a bit :)Kiana H. shared thisWe're excited to share our first blog post, introducing EnCompass: a framework that fundamentally rethinks how AI agents search and reason through complex, multi-step tasks. Presented at NeurIPS 2025. The Challenge: As AI agents tackle complex workflows like code translation or system debugging, possible paths multiply exponentially and small errors compound, yet traditional approaches hard-code search strategies directly into workflows, making them brittle and difficult to improve. Our Approach: EnCompass separates what agents do from how they explore solutions. Developers annotate workflows with branchpoints, then apply various search strategies such as beam search, best-of-N, and tree search without rewriting lines and lines of code. The Results: Across our three case studies (code translation, hypothesis generation, iterative code refinement): • 3-6× fewer lines of code to implement search strategies • Cleaner, more maintainable agent workflow code • Practical implementation of techniques to unlock better scaling laws Read the full blog post: https://lnkd.in/gu8cH72y Paper: https://lnkd.in/g-bCXCtz
-
Kiana H. reposted thisKiana H. reposted thisIf you're looking for the next ambitious project to take on, we’d love to talk! We’re building out the AI infra and compute to help people invent the systems of the future. Come build with us → hello@asari.ai, open jobs at jobs.ashbyhq.com/asari.ai
-
Kiana H. liked thisI think there is a certain kind of recursive delight that craftspeople of all generations intuitively understand. Let's call it a toolmaker's delight: a tickle, a nod, an invitation into the elegant world of a craft within their craft. It occurs when a woodworker builds the wooden workbench on which all their future woodworking will be possible, when a computer scientist writes the compiler for a language in that language, when a playwright stages a play inside their play, or when a dad reaches for a wrench to tighten the rollers on his toolbox. I was able to access this ancient delight by helping build a self-improving machine that makes the machines people use substantially more efficient. Ours uses language models to improve language model inference, by +16% in some scenarios. This result has huge implications for the AI industry, because every percent improvement in language model serving throughput means the industry needs that much less hardware (or can gain that much more extra capacity). What excites me most is the formal rigorous core that makes our machine general. Like a woodworker itching to build beautiful furniture, I am eager to apply it to even bigger problems. For more details, see: https://lnkd.in/epEcwSCU https://lnkd.in/eRaVbXKcKiana H. liked thisOur co-inventor agents optimized the full AI inference stack, verifying every change with rigorous correctness checks. They worked autonomously on vLLM serving two of the largest open-source LLMs, DeepSeek v4 Pro and Z.ai GLM 5.2, on NVIDIA B200s. The agents improved throughput and interactivity by up to 16% across concurrency levels, running for about a day per level. Production inference performance depends on millions of lines of code, hundreds of configuration parameters, and runtime dynamics that shift with every workload. A fast kernel in isolation doesn’t translate into a faster system overall. Our agents optimize everything together: kernels, schedulers, load balancers, and configuration. Speed only counts if model behavior is preserved. Customizations can quietly alter outputs while still acing standard benchmarks. We check every change with statistical distribution-matching tests on token-level probabilities, across a diverse multilingual corpus. The agents also learn as they go. A lesson from optimizing DeepSeek v4 Pro on how to avoid a class of distributed deadlocks transferred directly to GLM 5.2, saving 44 minutes of wasted wait time. Each run makes the next one better and faster. Full write-up at https://lnkd.in/gnVupZ35.
-
Kiana H. liked thisKiana H. liked thisOur co-inventor agents optimized the full AI inference stack, verifying every change with rigorous correctness checks. They worked autonomously on vLLM serving two of the largest open-source LLMs, DeepSeek v4 Pro and Z.ai GLM 5.2, on NVIDIA B200s. The agents improved throughput and interactivity by up to 16% across concurrency levels, running for about a day per level. Production inference performance depends on millions of lines of code, hundreds of configuration parameters, and runtime dynamics that shift with every workload. A fast kernel in isolation doesn’t translate into a faster system overall. Our agents optimize everything together: kernels, schedulers, load balancers, and configuration. Speed only counts if model behavior is preserved. Customizations can quietly alter outputs while still acing standard benchmarks. We check every change with statistical distribution-matching tests on token-level probabilities, across a diverse multilingual corpus. The agents also learn as they go. A lesson from optimizing DeepSeek v4 Pro on how to avoid a class of distributed deadlocks transferred directly to GLM 5.2, saving 44 minutes of wasted wait time. Each run makes the next one better and faster. Full write-up at https://lnkd.in/gnVupZ35.
-
Kiana H. liked thisThe next phase of AI won't just be more capable models. It will be AI systems that improve other AI systems. Asari AI's latest work is an example of this shift: autonomous agents that optimize the AI inference stack, improving performance while verifying correctness at every step. Over time, breakthroughs like this will make AI more useful, more accessible, and better equipped to help people solve some of the world's most challening problems.Kiana H. liked thisOur co-inventor agents optimized the full AI inference stack, verifying every change with rigorous correctness checks. They worked autonomously on vLLM serving two of the largest open-source LLMs, DeepSeek v4 Pro and Z.ai GLM 5.2, on NVIDIA B200s. The agents improved throughput and interactivity by up to 16% across concurrency levels, running for about a day per level. Production inference performance depends on millions of lines of code, hundreds of configuration parameters, and runtime dynamics that shift with every workload. A fast kernel in isolation doesn’t translate into a faster system overall. Our agents optimize everything together: kernels, schedulers, load balancers, and configuration. Speed only counts if model behavior is preserved. Customizations can quietly alter outputs while still acing standard benchmarks. We check every change with statistical distribution-matching tests on token-level probabilities, across a diverse multilingual corpus. The agents also learn as they go. A lesson from optimizing DeepSeek v4 Pro on how to avoid a class of distributed deadlocks transferred directly to GLM 5.2, saving 44 minutes of wasted wait time. Each run makes the next one better and faster. Full write-up at https://lnkd.in/gnVupZ35.
-
Kiana H. reacted on thisKiana H. reacted on thisI've taken the leap! I'm thrilled to share that I've accepted a full-time role as a Bedrock Rotational Analyst with Fervo Energy, starting in Houston! 🎉 The program is well designed to give experience in four business divisions through a 6-month rotation in each, and I am so honored to have been accepted out of the hundreds who applied. I have already met so many wonderful people at Fervo and am excited to kick off the program working with Sarah Jewett in strategy. Fervo is doing groundbreaking work in geothermal energy (#FervoEnergy), and if you've followed their journey at all, you know exactly why I'm so excited. I can't wait to contribute to the future of energy. Here's to the next chapter! 🌋
-
Kiana H. reacted on thisKiana H. reacted on thisToday I went to the GPT-5.5 launch party at OpenAI and had a chance to meet Sam Altman! One thing being San Francisco taught me is my perspective on reality. To achieve what you want, you usually need two things: an enabler + action The enabler creates possibility. Your action is what makes it happen. You are more in charge of your reality than you think. The acceptance rate for the event was apparently around 1.2%. Looking around the room, there were maybe ~150 people there, which means well over 10K people probably signed up. 5 years ago I would’ve never imagined this would be my reality. But here I am.. I wrote a lot about OpenAI ChatGPT on my personal blogs a few years back in the early days of GPT :) What a full circle that I just went to the GPT 5.5 Launch event!
-
Kiana H. reacted on thisKiana H. reacted on thisYesterday we introduced Fin for Sales, the first new role for Fin, our Customer Agent 🚀 The core idea is pretty simple. One Customer Agent that can take on different roles across the entire lifecycle, from first touch to years in. In a sales role, that means engaging at the right moment, helping people find the right fit, and qualifying conversations in real time so good opportunities don’t slip through. No handovers and no fragmentation. Just one continuous experience, the way it should be!
-
Kiana H. liked thisKiana H. liked thisBuilding software for space is one of the hardest engineering problems there is. As AI becomes more capable of writing and reasoning about complex code, we wanted to explore the key question: can AI help us invent at the frontier? Here's what we found →
Experience & Education
Volunteer Experience
Projects
-
Ares, Mars, and the in Between: The Literary Development of a War God
-
B.A. Honors Thesis in the Classics Department
Abstract: Ares-Mars in the modern conception is overwhelmingly that of unbridled violence, as evidenced not only by the character in Wonder Woman and the DC comic series but in numerous other contemporary depictions: in Rick Riordan’s Percy Jackson and the Lightning Thief as the primary antagonist, as the predominant rival character in the video game God of War, and again as an adversary in the Netflix series Blood of Zeus. In all of these…B.A. Honors Thesis in the Classics Department
Abstract: Ares-Mars in the modern conception is overwhelmingly that of unbridled violence, as evidenced not only by the character in Wonder Woman and the DC comic series but in numerous other contemporary depictions: in Rick Riordan’s Percy Jackson and the Lightning Thief as the primary antagonist, as the predominant rival character in the video game God of War, and again as an adversary in the Netflix series Blood of Zeus. In all of these forms of media, Ares-Mars and his attributes of rage and violence are utilized to oppose the interests of the main character. However, the preeminent role of Mars in Roman civilization is as the father of Romulus and Remus, and by extension the pater patriae, an unquestionably respected and worshipped leader. This dichotomy raises the question of the inconsistency between the Greek and Roman equivalents, as well as the discontinuity from the ancient depiction of the god with his modern portrayal. At some point in the process of syncretism, there was a reason that the Greek Ares became unified with the Roman Mars. However, the multitude of deities in the Greco-Roman world with dominion over war in addition to the many dimensions of both Ares and Mars inevitably creates complications in our perception of what ultimately gets passed down as a single, overwhelming figure of war. In particular, this thesis aims to understand how and why one aspect of Ares-Mars—namely the concept of wanton violence—becomes focalized and dominates the Western canon, as well as uncovering the unique strands of his identity as a religious figure and literary character.
Honors & Awards
-
Phi Beta Kappa
-
Elected during coterm year for exceptional academic achievement and commitment to the liberal arts and sciences. Recognized as one of the oldest and most prestigious academic honor societies in the United States.
-
Senior Award
Stanford Classics Department
-
Junior Award
Stanford Classics Department
-
Latin Department Award
-
-
California Junior Classical League Senior Scholarship
California Junior Classical League
Languages
-
Mandarin
Native or bilingual proficiency
-
Latin
Limited working proficiency
-
Greek, Ancient (to 1453)
Limited working proficiency
-
English
Native or bilingual proficiency
Organizations
-
DR Ultimate
Co-Captain (2023, 2024), Social Media Manager (2021-present)
- Present• Strategize recruitment/tryout process, practice plans, team-wide social events and delegate other team duties • Creator and manager of team Instagram @dr_ulti; designed all original posts using Procreate
-
Stanford O-Tone A Cappella
Vice President (2021-23), Financial Officer (2020-21)
-• Conceptualized social media posts, publicity flyers, costumes, club merch, and audition materials • Coordinated rehearsals and performances, facilitate intra-group communications, lead soprano sectional rehearsals • Budgeted annual club expense of $20K as Financial Officer (2020-21) and led financial grant application writing
-
Stanford Women's Ultimate
Co-Captain (2021-22), Tournament Director (2023-24)
-• Managed bi-monthly tournament logistics, expanded team bonding events and established a closer community • Independently designed season uniforms, created publicity material, started and ran the team Instagram • Initiated planning of the inaugural end-of-year banquet and organized venue, catering, and decorations
-
Stanford Calypso
Member
-• Member in Stanford's performing steel pan ensemble
-
California Junior Classical League
Co-Convention President
-• Planned an 2-day event of 1,000 people and a budget of $100K
Explore more posts
-
Nathaniel Miller, PhD
If you have a PhD, is now the absolute worst time to go into industry? I’ve seen some discussion about this recently and thought I would throw in my two cents: maybe, but there’s never a bad time to at least position yourself to pivot. There’s a lot of discourse about whether AI is slaughtering the market for entry-level roles. On the one hand, some good studies have shown that this is happening, but on the other, macroeconomic factors like tariffs and rightsizing after the pandemic-induced hiring splurge are also holding back hiring. Whatever the reason, things are tough. I recently decided to try out LinkedIn Premium — one of the insights it's provided is that for some of these roles that get 200 applicants in the first day, 10%–20% of the applicants have PhDs. I’ve heard anecdotes to the same effect. All this is exacerbated by the overproduction of PhDs for years and the subsequent collapse of the academic job market in many fields, although that goes back to the 2007–8 financial crisis. Anyway, I left academia in the summer of 2023, and it wasn’t easy then. It took me hundreds of applications (following on working with a career consultant, networking, and tailoring the bejesus out of every resume) to land a single offer. The fact of the matter is, even if PhDs have transferable skills and a lot of overall competence to offer, sometimes it takes a couple of years to build up some marketable subject matter expertise, and for employers to see your commitment to your career change. My pivot was not like most of the ones I heard about in 2021–22; I didn’t walk into a six-figure job after a few dozen applications that resulted in competing offers. But I’m glad I left when I did, even if it wasn’t easy, because now I have two solid years of experience in a couple of different spaces (first compliance and then tech). If I had waited until say, this year, it would have been even worse! But if you are thinking of trying to pivot from academia now, I don’t think the right attitude is to keep taking shelter in post-docs or adjunct gigs, hoping that the job market will improve. PhDs are always going to have to pitch their skills to a skeptical audience, and that’s why I said at the beginning, there’s never a bad time to start positioning for a pivot. Positioning could mean talking to a career coach, networking, upskilling, getting a part-time or contract role, or just getting out there and cold applying. Yes, except for AI, the job market sucks right now, but chances are pretty good it’s going to take you a while whenever you start.
41 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content