Skip to content

Latest commit

 

History

History
131 lines (75 loc) · 8.34 KB

File metadata and controls

131 lines (75 loc) · 8.34 KB

Kartik Jevaji

Full Stack AI Engineer · M.A. Statistics · St. Paul, MN (412) 638-3064 · kartikinpublic@gmail.com nitrotap.dev · github.com/nitrotap · linkedin.com/in/kjevaji · huggingface.co/nitrotap

Software engineer building AI systems and the machinery that establishes whether they work. Ontologies and automated-reasoning agents that keep language models honest; evaluation harnesses with graders where proof does not apply; formal verification where it does. Master's in statistics, an enterprise implementation background, and full ownership from schema through interface to production deploy.


Experience

Phase Change Software — Software Engineer

Nov 2023 – Present · Hybrid

  • Building a neurosymbolic AI agent with autonomous A2UI integration.
  • Introduced harness-level evaluations and graders for accuracy and hallucination detection across agentic applications and MCP servers, integrated with LangSmith. Adopted internally by QA.
  • Developed agentic and domain ontologies.
  • Fine-tuned small language models (Qwen, Llama 3.2, Granite, DeepSeek) for COBOL system understanding using Python and HuggingFace Transformers.
  • Own the CI/CD pipeline: weekly production deployments gated by 300+ Playwright end-to-end tests, with code review and production support.
  • Migrated the company website from WordPress to React/Next.js, raising Lighthouse from 60–80% to 99–100% and reaching top-10 search rankings.
  • Built a ReasonReact component library and established the team's unit testing framework; developed Angular prototypes with embedded agent features.

edX — Instructional Specialist

Mar 2023 – Dec 2023 · Contract, Remote

  • Supported a full-stack web development classroom: tracked metrics, led discussions, and troubleshot student issues to drive high completion rates.

ATHENA Consulting — Web Development Intern

Jul 2023 – Nov 2023 · Part-time · Golden, CO

  • Designed and delivered a client website on schedule, applying responsive design to resolve mobile usability issues.

Epic — Application Coordinator, Implementation Services

Apr 2011 – Jan 2013 · Verona, WI

  • Led the Epic implementation in Omaha, NE: coordinated the rollout across five analysts, owned day-to-day troubleshooting and task management for the build team, and served as primary point of contact for customer questions.

Alzheimer's Association — Volunteer, Data Entry / Tech Support

2017 – 2022 · Denver, CO

  • Maintained accuracy of key database identifiers; provided on-demand technical support during education classes; automated processes with the Google Sheets API.

Projects

stats-claw — Rust statistical computing

crates.io · docs.rs · GitHub

In-process statistical computing with zero runtime dependencies — distributions, hypothesis tests, and resampling, validated against scipy.stats, scikit-learn, ruptures, and mlxtend through committed golden fixtures to documented tolerances between 1e-8 and 1e-12.

68 Kani proof harnesses in the published crate (72 across the source workspace), each with its bound stated: permutation bijectivity, k-fold partitioning, Welford variance non-negativity, PRNG unit-interval bounds, ziggurat table indices. Miri as the undefined-behaviour backstop. Fourteen distribution families; exact and asymptotic p-values across t, ANOVA, χ², Fisher, Mann–Whitney, Wilcoxon, Kruskal–Wallis, KS, and Shapiro.

Verus — merged compiler contribution

PR #2235 · write-up

Fixed type arguments being dropped during Ghost/Tracked macro expansion in the Verus verifier, so a turbofish written in source survived to the type checker. Resolves issue #2013, filed by a project maintainer. Merged March 2026.

yee-claw — ontology-driven verified code generation

Private. A TypeDB ontology is the single source of truth; a zero-dependency generator compiles it into the Rust type layer that hand-written numerics are built against. Abstract entities become marker traits with the hierarchy preserved; concrete entities become structs carrying every inherited attribute. stats-claw is a mechanical carve-out, produced by a configurable tool rather than copied.

cobol-coder — small-model fine-tune

HuggingFace

Qwen2 0.5B fine-tuned for COBOL system understanding with Unsloth and HuggingFace Transformers. Part of a broader effort on legacy mainframe code, where training data is scarce and wrong answers are expensive.

COBOLEval — code-generation benchmark

GitHub · built on bloop.ai's benchmark

Substantial rebuild of bloop's HumanEval-to-COBOL transpilation (146 tasks; dataset unchanged). Added a JCL parser, validator, and linter for mainframe job control (static analysis only), a Rig-based agentic evaluation orchestrator in Rust with per-task JSONL conversation logging, Ollama support for local inference, and GnuCOBOL compatibility fixes.

Telesto platform

Private. Federated multi-tenant platform serving several brands from one codebase — contract-first internal packages, ontology-backed data layer, Ansible-managed VPS deployment with nginx and containerised services. Two customer engagements.

Proximal — training-data pipeline

Private. Continuously running pipeline streaming merged pull requests from the public GitHub Archive, enriched with diffs, reviews, and commits, exported as JSONL training sets with language-specific subsets refreshed on schedule. Deduplicated on event ID; no API cost.

Earlier

  • City Code Assistant — retrieval over municipal code so a resident can find a permit checklist without reading the ordinance. GitHub
  • museum-models — 3D PWA in Ionic React, Three.js
  • electron-kiosk — cross-platform Electron shell wrapping Three.js
  • Mental Health Check — mood-tracking PWA. MERN, Apollo/GraphQL, Chart.js
  • Brain Lift — NASA Task Load Index cognitive-load tracker across web, iOS, Android. LAMP, Angular, Ionic, Capacitor

Technical

Verification & correctness — Kani (bounded model checking), Verus (linear ghost types), Miri, property and invariant design, reference-equivalence testing, deterministic PRNG design, Playwright, Jest/Vitest

AI engineering — LLM fine-tuning (Unsloth, HuggingFace Transformers; Qwen, Llama 3.2, Granite, DeepSeek), OpenAI Agents SDK, LangChain/LangSmith, MCP, multi-agent architectures, harness-level evals and graders, hallucination detection, RAG and GraphRAG, vector stores, NeMo Guardrails, Azure OpenAI, Anthropic Claude, OpenAI

Symbolic AI & knowledge representation — TypeDB/TypeQL, agentic and domain ontologies, ontology-driven code generation, LinkML, Neo4j, semantic web (Jena, RDF), Datalog

Languages — Rust, TypeScript/JavaScript, Python, Kotlin, SQL, COBOL/JCL, PHP, Java, R, SAS, OCaml/ReasonML

Web & product — React, Next.js, TanStack Start, Angular, Ionic/Capacitor, Three.js, React Three Fiber, Astryx, Radix UI, Tailwind, Drizzle ORM, better-auth, GraphQL/Apollo, Express, Ktor

Data — PostgreSQL, TypeDB, MySQL/MariaDB, MongoDB, pgvector, Neo4j

Infrastructure — Docker, Ansible, nginx, systemd, VPS administration, AWS (EC2, Lightsail, Route 53), Azure, Vercel, CI/CD, Turborepo/pnpm, Sentry

Statistics — hypothesis testing, Bayesian statistics, resampling and bootstrap, regression and regularisation, clustering and decomposition, change-point detection, time series

Design — Figma, Sketch, wireframing, design systems, style guides, accessibility (WCAG AA)


Education & Certifications

CompTIA A+ Certification Feb 2024 – Feb 2027
Regis University Certificate, Web Development — LAMP Aug 2022 – Jun 2023
University of Minnesota Certificate, Full Stack Web Development — MERN Dec 2021 – Jun 2022
University of Pittsburgh M.A. Statistics Aug 2009 – Aug 2010
University of Pittsburgh B.S. Statistics 2006 – 2009