Work

Projects

Each card carries the problem it solved, the figures it moved, and what I owned on it. Every number was measured in the source repository.

  • 2025 – 2026

    Agentic Development Board

    Turning a business request into shipped code meant a chain of manual handoffs — speccing, decomposition, implementation, tests, documentation — with no single place to answer where a task actually is.

    Feature branches merged
    94Feature branches merged
    Workflow types
    16Workflow types
    MySQL schema
    46 tablesMySQL schema
    Project size
    ~116K LOCProject size

    What I owned — Initiated the platform and led it as maintainer; personally built the LLM provider abstraction, the subprocess layer, the Atlassian auth layer and the self-healing API-generation workflow.

    • Python
    • FastAPI
    • LangGraph
    • Kafka (KRaft)
    • MySQL
    Repo 
  • 2025 – 2026

    AI Agent Kits & Spec Skills

    Coding agents produce plausible code that ignores a codebase's real conventions, and every team member was re-teaching the same context to the same models by hand.

    Agent definitions
    16Agent definitions
    Lines, sole author
    38KLines, sole author
    Blame-verified
    100%Blame-verified

    What I owned — Sole author of all 16 agent definitions and the conventions behind them.

    • Claude Code SDK
    • Cursor CLI
    • MCP
    • Python
    • Markdown
    Repo 
  • 2024 – 2025

    Enterprise Knowledge RAG

    Business-flow knowledge lived in Jira and Confluence, so answering a routine process question meant interrupting an engineer who happened to remember the ticket.

    Blind pairwise, LLM judge
    240–30–13Blind pairwise, LLM judge
    Human expert judge
    246–34–3Human expert judge
    Judge agreement
    91.5%Judge agreement
    Jira issues indexed
    30–60KJira issues indexed

    What I owned — Built the retrieval pipeline, the citation construction, the async serving layer and the blind evaluation that cleared it for production.

    • Python
    • LangChain
    • LangGraph
    • Pinecone
    • Flask (async)
    Repo 
  • 2025 – 2026

    CodeHarbor

    Reviewing every merge diff across an ERP estate for security defects is work no human team sustains, and a plain LLM reviewer produces confident findings it cannot substantiate.

    Commits
    207Commits
    Insertions
    ~25KInsertions
    Security checks
    9Security checks

    What I owned — Initiated the repo and led the scan engine — sandboxed execution, the 9-check checklist, evidence gating and de-duplication.

    • Python
    • FastAPI
    • LangGraph
    • Claude Code SDK
    • Pinecone
    Repo 
  • 2025

    LLM Evaluation Framework

    Deciding whether a cheaper model can replace an incumbent in production is usually settled by a demo, a benchmark table, or the price sheet — none of which replay what users actually send.

    Judge-choice swing
    37.4 pointsJudge-choice swing
    Verdicts failing to parse
    37.6%Verdicts failing to parse
    Cost of the candidate
    ~12× cheaper / tokenCost of the candidate

    What I owned — Designed and built the harness, the seven evaluation arms, the reciprocal cross-judging design and the self-preference probe.

    • Python
    • OpenAI
    • Anthropic
    • Gemini
    • DeepSeek
    Repo 
  • 2026

    protoframe

    Design-to-code tools emit markup that looks right in a screenshot and fails in a real application, and there is rarely any way to tell whether one generated design is better than another.

    Repository
    Initiated & ledRepository
    CSS policy
    Deny-by-defaultCSS policy
    Design quality
    Offline A/B harnessDesign quality

    What I owned — Wrote the root commit and led the build — generation pipeline, CSS allow-list and the offline evaluation harness.

    • Python
    • FastAPI
    • Celery
    • LangGraph
    • Angular 21
    Repo 
  • 2025

    n8n WhatsApp Agent Platform

    A WhatsApp support agent has to hold conversation state across bursty, out-of-order messages — and nobody knew what throughput the orchestration layer could actually sustain.

    Workflows / nodes
    22 / 357Workflows / nodes
    Sustained ceiling
    ~65 req/sSustained ceiling
    Load-test requests
    30,300 (26→940/s)Load-test requests
    Fleet coverage
    9 instances / 5,558 workflowsFleet coverage

    What I owned — Authored all 22 workflows and 357 nodes, the conversation-state design, the stale-reply guard, the load test and the fleet tooling.

    • n8n
    • Twilio
    • Redis
    • LLM tool calling
    • Python
  • 2024 – 2025

    Passport OCR Quality Pipeline

    Passport photographs arrive skewed, low-contrast and noisy, so OCR either fails or — worse — returns a plausible wrong field that a human has to catch.

    MRZ validation
    Check-digit verifiedMRZ validation
    Enhancers benchmarked
    DE-GAN / GFPGAN / NAFNetEnhancers benchmarked

    What I owned — Built the restoration pipeline, the MRZ validation path, the deep-enhancer benchmark and the automated QA gate.

    • Python
    • OpenCV
    • EasyOCR
    • PassportEye
    • NumPy
    Repo