Kayvan Zahiri
Kayvan Zahiri

Kayvan Zahiri

AI engineer in San Francisco. Founder of BuildGist and ResumeAI.

Email me Resume
ParkCast SFParking predictions for 12.7K SF blocks ATS share, from ResumeAI data Merged openai/whisper #2836 Fix English number normalizer

Work

BuildGist run page explaining an added Google sign-in, with a high-severity concern that any signed-in user can delete anyone else's note
BuildGist app icon

89 ms

on the agent’s critical path

100 files
149 ms

Founder. macOS app.

BuildGist

Explains what Claude Code, Cursor or Codex changed after every run that matters. A local classifier picks those runs: a one-line change to an auth file outranks a 300-line README rewrite. The model has to cite an exact quote for every claim, and each citation is checked in code; a comment or a test’s name can show what someone meant, but only executable code can establish what the code does. Releases are gated on an eval harness for factual accuracy, hallucinations and planted-bug detection.

Python, Rust (Tauri), Claude API, Stripe, Fly.io

buildgist.com

ResumeAI editor with a live resume preview
Example ResumeAI score card reading 92
Bar chart of applicant tracking system share among large employers ATS share, from ResumeAI data

Founder. Built solo.

ResumeAI

AI-powered ATS resume optimizer: ATS compatibility scoring, Claude-powered bullet rewriting, AI cover letter generation, PDF/DOCX export and a companion Chrome extension. Reached active users across 7+ countries with 80%+ week-over-week new-user growth.

Next.js, TypeScript, Supabase, Claude API, Stripe, Chrome extension

withresumeai.com

ParkCast SFPredicted occupancy around Chase Center

Team project

ParkCast SF

End-to-end MLOps system predicting parking occupancy for 12.7K SF blocks at 8.98 MAE / 0.73 R² from a 33-feature LightGBM residual model. Fully automated GitHub Actions retraining pipeline with MLflow tracking and a champion-challenger gate that blocks silent regressions before Cloud Run rollout.

FastAPI, LightGBM, Docker, GCP Cloud Run, GitHub Actions, MLflow

Live demo GitHub

PantrIQ Discover screen with featured recipes
PantrIQ meal planner week view
PantrIQ app icon

iOS app

PantrIQ

Full-stack AI mobile app for pantry management, with recipe suggestions, automated meal planning and shopping lists from scan, voice or text input.

React Native, TypeScript, Supabase, OpenAI, TailwindCSS

App Store

Word error rate, Whisper

Twenties6.53% Seventies4.67%

Utterances with a 700 ms pause

Twenties8.0% Sixties19.7%
2,760 Common Voice clips matched on accent, gender and speaker

Benchmark

asr-age-gap

A published benchmark that set out to find an age penalty in speech recognition and found the opposite. On 2,760 Common Voice clips matched on accent, gender and speaker, Whisper scores 4.67% WER on people in their seventies against 6.53% on twenties, replicated on a wav2vec2 CTC control. The penalty is in turn-taking instead: at a fixed 700ms endpoint, 19.7% of sixties utterances carry a pause that long against 8.0% of twenties, which word error rate cannot see.

PyTorch, Whisper, wav2vec2, Common Voice

GitHub

Meridian Core, a legacy banking test app, on its Confirm Funds Transfer screen

Agents and automation

agent-hands

Computer-use automation for legacy back-office apps with no API. A model drives the app once and emits a reviewable, versioned capability artifact; a deterministic engine replays it with no model in the loop. Targets are recorded as a ranked list of accessibility-tree strategies, so a renamed field degrades to a fallback instead of breaking, and the degradation is logged. Built against a deliberately hostile fixture: frameset, table layout, no labels, server-generated IDs.

Python, Playwright, Claude API, Accessibility tree

GitHub

71/75

answered or refused correctly

False answers
0
Against a no-data LLM
23–3

Evaluation

EARTHBENCH

A grounding harness that tests whether an agent-facing geospatial API can be trusted, scored against oracles outside the system under test. It answers or refuses correctly on 71 of 75 cases with zero false answers, and beats a no-data LLM 23–3. The finding that mattered was reproducibility: the same question at the same coordinate returned different field selections across calls, which for a product sold as audit-ready is the result worth reporting.

Python, Eval harness, Bootstrapping

GitHub

More projects

  • handoff-note

    A Claude Code plugin for switching agents. /handoff writes a HANDOFF.md that quotes your request exactly as you typed it, then lists the decisions made, what changed, what’s left and how to check the work, so Codex, Cursor or a fresh session can pick up without you explaining it again.

    Claude Code plugin, Skills

    GitHub

  • brainwire

    Lossless codec for Neuralink N1 electrode recordings: 2.79× across all 743 files, ahead of xz -9e while encoding 4.6× faster with 49 KB of model state. The design came from measuring the data first: the samples sit on a 64-step ADC lattice, which makes plain first difference the best predictor.

    C, Entropy coding, Compression

    GitHub

  • State of ATS 2026

    Open dataset of which applicant tracking system 738 large employers use, with a keyless npm API. Version 2.1.0 re-checked the snapshot against live careers portals and cut the verified count from 704 to 421.

    Dataset, Node.js, npm, MIT

    GitHub

  • Search Engine

    Full-stack search engine with inverted index, TF-IDF ranking and a multithreaded web crawler. 1,746 SLOC with 1,910 lines of Javadoc documentation.

    Java 21, Apache OpenNLP, Jetty, Bootstrap, Maven

    GitHub

  • Semiparametric Regression Visualizer

    Interactive web app for exploring semiparametric regression models. Compare OLS against Generalized Additive Models with real-time visualizations on datasets like Boston Housing and World Happiness.

    Python, Streamlit, pygam, statsmodels, Google Cloud Run

    GitHub

  • AI Interviewer Tool

    GenAI research app that simulates user interviews and focus groups with participant personas you define, for product and UX research.

    Python, Streamlit, OpenAI, Prompt engineering

  • AI/ML Industry Data Dashboard

    Data dashboard aggregating AI/ML content from arXiv papers and Medium articles, with automated scraping via GCS, article trend visualizations and Anthropic-powered topic clustering.

    Python, Streamlit, GCS, NLP, Anthropic, Google Cloud Run

  • PeopleCodeOpenAI

    Python library that gives easy access to OpenAI’s tools and APIs, including text-to-speech, speech recognition and RAG. Used in introductory courses at USF.

    Python, OpenAI API, RAG, Speech recognition

Open source

91 pull requests merged into third-party production repos, with more in review.

Merged openai/whisper #2836 The English number normalizer rewrote the digit 1 as “one” inside larger numbers, so timestamps and quantities came out wrong. Twelve lines. Merged by one of the authors of the Whisper paper.
Merged google/sentencepiece #1320 and #1352 On an empty batch, the return_bytes flag reached the binding as a list instead of a bool. Merged by Taku Kudo, who wrote SentencePiece.
Merged NVIDIA-NeMo/Speech #16201 The end-of-utterance evaluator could report negative false negatives, which are counts, after a prediction cut a reference short.
Merged microsoft/onnxscript #3097, #3098, #3099, #3101 and #3102 Five fixes to the graph rewriter, starting with Relu/Clip chains it fused even when their ranges didn’t overlap.
Merged IBM docling #4836 Whisper writes Chinese and Japanese with full-width 。？！, so a whole CJK transcript merged into one sentence and diarization gave it a single speaker.
Landed facebookresearch/faiss b4c66ba The index factory’s round-trip silently dropped the HNSW storage index. Imported and landed by a Meta engineer.
Merged NVIDIA/DALI #6471 Hue in ColorJitter came out 360× too small, so hue=0.5 rotated half a degree.
Merged astral-sh/uv #21144 and #21146 Two fixes into uv. Merged by Astral’s co-founder.
Merged librosa #2087 Input validation in rms. Merged by the library’s creator.
Merged elevenlabs-python #858 Closed as untouchable generated code until I showed the file was listed in their own .fernignore.
In review scikit-learn #34716 RadiusNeighborsRegressor silently returned a sentinel value for empty neighborhoods with integer targets.
github.com/Kayvan-Zahiri

Experience

  1. Oct 2026 – PresentRemote

    Clipboard Sports

    Software Engineer

    • Full-stack engineering for a sports technology startup, working directly with the founder and engineering team on the initial development roadmap across frontend, backend and product
    • Own debugging, testing and launch-readiness work on core product features

  2. Jul 2026 – PresentSan Francisco, CA

    University of San Francisco

    Teaching Assistant

    • Supported two graduate Data Science (MSDS) courses, with a third upcoming

  3. Jan 2026 – Jun 2026Remote

    Asurion

    AI Engineer

    • Built a voice turn-detection system, improving auto-labeling accuracy from 50% to 88%+ (90%+ on a 400-sample validation set)
    • Designed and scaled an automated labeling pipeline to auto-label 400K+ audio samples, enabling large-scale fine-tuning of a Whisper-based PyTorch model
    • Deployed the model in a real-time voice AI pipeline (LiveKit) for low-latency, end-to-end conversational agent testing

  4. Oct 2025 – Jan 2026Remote

    Spotly Jobs

    AI Engineer

    • Engineered ATS data pipelines using dlt and dbt, expanding job posting coverage by 5%+ and scaling automated processing to 500+ postings per pipeline run
    • Built 7 Dagster assets and sensors to automate ETL workflows, cutting manual engineering intervention by 50%+ and saving 10+ engineer-hours per week
    • Improved data ingestion throughput by 30–40% through optimized scraping logic and REST API handling, processing 150–200 additional postings per hour at no added infrastructure cost

  5. May 2024 – Sep 2025Remote

    Outlier AI (Scale AI)

    AI Engineer

    • Trained LLMs using RLHF, improving instruction-following accuracy by 20%
    • Evaluated 1M+ tokens across 5+ datasets, identifying 300+ instruction-following errors that directly informed targeted fine-tuning
    • Refactored 3+ AI models to generate more precise and succinct output

  6. Sep 2024 – May 2025San Francisco, CA

    University of San Francisco, IDEA Team

    Software Developer

    • Lead developer delivering 3 full-cycle applications for the IDEA team
    • Created 3 GenAI Python apps adopted by 50+ students and professors
    • Researched and evaluated 10+ existing AI tools and technologies

  7. Aug 2024 – Dec 2024San Francisco, CA

    Founders Network

    Software Developer

    • Built Slack integrations into a codebase of over 10,000 lines of code
    • Delivered 8 sprint cycles in Agile with a 95% on-time delivery rate
    • Collaborated directly with the CEO and senior engineering team, with executive-level product visibility

  8. May 2024 – Aug 2024San Francisco, CA

    University of San Francisco

    Software Developer

    • Created and deployed 2 generative AI Thunkable apps, one featured on the site’s front page
    • Lead developer on a full-stack PeopleMuseum project, 5,000+ lines of code
    • Built front-end UI with React JS, achieving load times under 2 seconds
Portrait of Kayvan Zahiri

About

I’m an AI engineer and a software engineer at Clipboard Sports, a sports technology startup. I founded BuildGist, a Mac app that explains what coding agents changed, and ResumeAI, an AI-powered resume platform I built solo that’s now used across 7+ countries. Before that I was on the Voice AI team at Asurion, where I built a voice turn-detection system and scaled an automated labeling pipeline to 400K+ audio samples, used to fine-tune a Whisper-based model for real-time conversational AI.

With a B.S. in Computer Science and an M.S. in Data Science & AI from the University of San Francisco, I work across the full AI/ML lifecycle, from training and fine-tuning LLMs with RLHF to shipping production MLOps pipelines and full-stack apps end to end.

AI / ML
LLM fine-tuning, RLHF, RAG, voice AI, prompt engineering and model evaluation
Full-stack
React, Next.js, TypeScript, Python. Apps from database to deployment
Cloud & MLOps
AWS and GCP certified. MLOps pipelines, ETL, Docker, distributed computing

Skills

Languages
Python, TypeScript, Java, C, SQL, NoSQL, JavaScript, HTML/CSS
AI / ML
Generative AI, LLM fine-tuning, RLHF, voice AI, prompt engineering, model evaluation, RAG, deep learning, PyTorch, Whisper, Claude API, OpenAI
Data science
PySpark, SparkSQL, Pandas, NumPy, Scikit-Learn, Matplotlib, Plotly, ETL pipelines, regression, classification, unsupervised learning
Tools & platforms
Claude Code, Cursor, Codex, AWS, GCP, Docker, Git / GitHub, Bash, LiveKit, MLflow, Dagster, Apache Airflow, dlt / dbt, MongoDB, distributed computing
Frameworks
React / React Native (Expo), Next.js, Tauri, FastAPI, Supabase, Stripe, TailwindCSS, Streamlit, Bootstrap, Eclipse Jetty, Maven

Education

  • M.S. in Data Science & Artificial IntelligenceUniversity of San FranciscoJul 2025 – Jun 2026
  • B.S. in Computer ScienceUniversity of San FranciscoAug 2021 – May 2025
  • AWS Certified Cloud PractitionerAmazon Web Services
  • Google Cloud Digital LeaderGoogle Cloud

Contact

kzahiri@dons.usfca.edu

The fastest way to reach me is email. I’m looking for AI engineering roles and interesting problems to work on.

LinkedIn GitHub Resume (PDF)