resume.yaml

profile:
  location: Davis, CA
  name: Devang Borkar
  phone: "+1 (312) 358-5722"
  email: devangborkar3@gmail.com
  links:
    - label: LinkedIn
      url: https://www.linkedin.com/in/devang-borkar-710b49201/
    - label: GitHub
      url: https://github.com/devangb3
    - label: Portfolio
      url: https://devang-borkar.netlify.app/

skills:
  - category: Languages
    items: Python, C#, Java, TypeScript/JavaScript, SQL, Go, C++
  - category: AI/ML
    items: LLMs, Multi-agent systems, LoRA fine-tuning, Prompt engineering, ReACT agents, tool-calling, RAG, embeddings, Harness Engineering, Synthetic Data Generation, Agentic Memory, FAISS
  - category: Frameworks
    items: PyTorch, HuggingFace Transformers, FastAPI, LangChain, React, MongoDB, Redis
  - category: Cloud/APIs
    items: CI/CD Pipelines, GCP, AWS, Docker, OpenAI/Anthropic/ Gemini APIs, OpenRouter
  - category: Coding Agents
    items: Claude Code, Codex CLI, Cursor IDE, OpenCode, Github Copilot

education:
  - degree: Masters of Science in Computer Science
    institution: University of California, Davis
    gpa: "GPA: 3.91/4.0"
    date: Sept 2024 - June 2026
    coursework: Distributed Databases, Machine Learning, Data Structures and Algorithms, Vision Language Models
  - degree: Bachelor of Engineering in Computer Engineering | Honors Track in AI/ML
    institution: Pune University, India
    gpa: "GPA: 4.0/4.0"
    date: Aug 2019 - June 2023

experience:
  - id: experience_pilotcrew_ai_engineer
    title: Software Engineer
    organization: PilotCrew AI
    dates: Oct 2025 -- Present
    location: Davis, CA
    bullets:
      - Developed end-to-end AI evaluation infrastructure for agentic systems, using Python APIs, MongoDB persistence, Dockerized execution, and observability workflows for repeatable model-quality measurement.
      - Built an automated prompt optimization pipeline that analyzes agentic traces and iteratively improve prompts, achieving a 6.63% performance lift for conversational agents on MultiChallenge benchmark.
      - Engineered agent harnesses for execution, tool-use tracing, hidden-state management, deterministic grading, and failure analysis across coding, interactive tool-calling and deep research workflows.
    supporting_projects:
      - id: pilotcrew-gen-eval
        title: Agentic Eval Platform
        url: https://pilotcrew.ai/autoeval/agentic-eval/about
        description: Built an evaluation platform for testing and improving LLM and agent reliability across structured tasks, coding agents, RAG workflows, tool use, and iterative prompt-optimization loops.
        technologies: Python, MongoDB, LLM Evaluations, Prompt Optimization, OpenRouter, Observability
        highlights:
          - Built repeatable evaluation infrastructure with structured run tracking, artifact export, failure-mode reporting, and observability views for comparing prompt and system changes.
          - Added runner support across coding, RAG, interactive tool-use, BrowseComp, MultiChallenge, tau2, and SWE-bench style experiments.
          - Implemented OpenRouter usage and cost tracking plus regression guards such as client-prompt acceptance gates, oracle gates, and gate-skip diagnostics.
  - id: experience_learnhaus_swe_intern
    title: Software Engineer Intern
    organization: LearnHaus AI
    dates: Jun -- Aug 2025
    location: Sacramento, CA
    bullets:
      - Built a full-stack multimodal AI application (React, Python) from scratch, processing video, audio, and text inputs with optimized latency for real-time user feedback.
      - Developed a LLM-as-Judge evaluation framework using GPT-5 and Gemini-3 to score model outputs across multiple quality dimensions; deployed to Google Cloud Platform(GCP)
    supporting_projects:
      - id: lh-multimodal-svc
        title: Emotion-Aware Feedback System for Public Speaking
        url: https://multimodal-svc-frontend-277660335430.us-central1.run.app/
        description: Built a multimodal service that processes video, audio, and transcripts to deliver presentation coaching, emotional analysis, and structured feedback for public speaking improvement.
        technologies: Python, Hume AI, Multimodal AI, Async Processing, Data Visualization
        highlights:
          - Engineered a pipeline for processing video, audio, and text signals into coaching-oriented feedback.
          - Integrated Hume AI for emotional analysis and built a visualization layer for actionable presentation insights.
          - Designed LLM-judge style quality control so one model could evaluate another model's responses.
  - id: experience_hexaview_software_engineer
    title: Software Engineer
    organization: Hexaview Technologies
    dates: Aug 2022 -- Sept 2024
    location: Pune, India
    bullets:
      - Shipped 20+ production features for a Fortune 500 fintech platform; contributed to scalable microservices using C# and SQL. Collaborated within a 30+ person cross-functional Agile team using JIRA.
      - Optimized a legacy .NET backend servicing over 1 million monthly requests, applying key design patterns to reduce code complexity & successfully redesigning 50+ REST APIs in MVC architecture

projects:
  - id: project_process_reward_model
    title: Process Reward Model for On-Device LLMs
    url: https://github.com/devangb3/Process-Reward-Models
    url_label: GitHub
    context: Graduate Research
    bullets:
      - Performed Supervised Fine Tuning on an 8B parameter LLM using PyTorch to verify reasoning steps; optimized for edge deployment with a lightweight 0.6B model.
      - Improved model accuracy by 21% on GSM8K using specualtive decoding over baseline methods; published model weights to HuggingFace.
  - id: project_resshare
    title: "ResShare: Decentralized File Sharing Platform"
    url: https://github.com/ResilientApp/ResShareDeployable
    url_label: GitHub
    context: Apache Foundation - ResilientDB
    bullets:
      - Built a full-stack file-sharing platform using Flask and React, implementing REST APIs for user authentication, access control, file upload/download, folder hierarchy management and sharing workflows.
      - Added a document indexing and RAG pipeline for files using text extraction, semantic chunking, embeddings, FAISS search, source-attributed responses, and per-user retrieval isolation.
  - id: project_causalflow
    title: "CausalFlow: Agentic Trace Debugging Framework"
    url: https://github.com/devangb3/CausalFlow
    url_label: GitHub
    context: NeurIPS '26 submission
    bullets:
      - Built a Python framework for causal attribution in multi-step LLM agent traces, identifying failure-inducing reasoning/tool steps and generating localized repair pairs that are useful as DPO post-training data.
      - Evaluated across 3,000+ math, code generation, web QA, and medical browsing tasks; repaired 42.7% of failed agent outputs and improved post-repair accuracy by a median +16.9 % over baseline.