Sorin Decu

I build AI workflows for regulated work.

Thirteen years in banking taught me how rule-bound work runs. I build the tools that do it faster and still pass review. Three demos are running on this page right now. Try to break them.

Live state of the three demos below. Change one and this changes.

plot: drawn in your browser

seed
0000 0000
curve
r = R + a*(sin(n*t + p) + h*sin(m*n*t + 2*p))
bands
n, k
drawn
0 line segments
sweep
000 of 360 degrees
frame rate
measuring

#statement-to-pnl

Statement-to-P&L pipeline

Turns a bank statement into a finished profit-and-loss statement for a small accounting firm.

problemFour associates covered about 100 client businesses. Every month they sorted roughly 150 bank statements into ledger codes by hand, at 10 to 15 minutes a statement.

buildAn associate emails a statement with the client’s industry in the subject line. The system parses it, extracts each transaction, assigns the ledger code that fits the industry, and generates the P&L as a PDF.

result
1–3 minper statement
before
10–15 min by hand
volume
about 150 statements a month
input
bank statement, by email
coding
language model, by client industry
output
P&L as a PDF
kind
client work, 2026
code
private

demo: statement to P&L

Pick a sample statement and run it.

Bank statement

March, made-up business

    Profit and loss

    March, built from the statement

    The same hardware store is materials for a landscaper and repairs for a restaurant, which is why the industry travels with the statement. This demo uses made-up data. The real pipeline uses a language model to choose each ledger code; here the codes are fixed so it can run in your browser.

    #software-factory

    Software factory

    A build pipeline where AI coding agents write the code, and tests and a person decide what ships.

    problemAI coding agents write code fast. They can also report success on work that never landed. I wanted every change checked the same way before it could merge.

    buildI file a ticket by chat. A router classifies it and hands it to a coding agent, which builds in its own isolated copy of the repository. The change has to pass automated gates and then my review. Only a completed merge can mark it shipped, and every step is written to an audit trail.

    coverage
    240+pytest tests on the factory’s own code
    intake
    ticket filed by chat
    builder
    AI coding agent, isolated git worktree
    gates
    tests, types, changes
    approval
    a person, on every merge
    shipped
    written only by a completed merge
    record
    audit trail of every step
    kind
    internal system, 2026
    code
    private

    demo: software factory

    A change is waiting for review. In this demo the reviewer is you.

    1. Ticket
    2. Router
    3. Build
    4. Gates
    5. Review
    6. Merge

    Fix the date format on the report page

    In review

    The agent’s change passed every gate. It merges only if a person approves it.

    3 files changed, 42 lines added, 7 removed

    audit trail

      A simulation of the pipeline’s stages. It isn’t connected to the real factory.

      #agent-fleet

      Supervised agent fleet

      Three AI agents in production since April 2026, with a health check that restarts them when they fail.

      problemOne agent sat silent for eight days while its health check stayed green. The check confirmed that a job was running and never asked whether the agent could still answer.

      buildI rebuilt the check to confirm three things: the process is alive, it can still be reached, and it is making progress. When a check fails it alerts me, and it restarts an agent that has died or hung. Then I broke the agents on purpose, four different ways, and fixed each gap the tests exposed.

      kill tests
      4 modescrash, dropped connection, reboot, hang
      agents
      3, in production
      since
      April 2026
      checks
      process, reachability, progress
      on failure
      alert, then restart
      kind
      operations, 2026

      demo: supervised agents

      Break an agent and watch the check catch it. Then switch to the old check and break one again.

      next check

      what the health check reported

        A simulation with time sped up. The real check runs every few minutes.

        #vibe-shipped

        Vibe Shipped

        A sourced directory of real apps built with AI coding tools, made for people and for AI agents to read.

        problemAnyone choosing an AI build tool, and any agent helping them, needs facts they can check: what was built, with which tool, and what it earned.

        buildEach entry summarizes a public post and links to it, and revenue claims are labeled by how they were verified. The same data is served through the site, an API, an open dataset and an MCP server, so an agent can ask it directly.

        tracked
        908MCP servers, probed every night
        tools
        7,878 indexed
        apps
        181, each with a public source
        probes
        26 nights in a row
        serves
        site, API, open dataset, MCP server
        listed in
        the official MCP Registry
        as of
        November 21, 2026
        kind
        live product, 2026

        captured: live MCP responses

        Ask it the way an agent would.

        Real responses, captured from the live MCP server on November 21, 2026 and shortened. The app count is from November 18.

        #ga-law-rag

        Georgia closing-law retrieval

        Answers questions about Georgia real estate closing law from the statute text, and cites every section it uses.

        problemA legal answer is only useful if the reader can check it. A model that knows the law but can’t show its source has to be taken on trust.

        buildStatute sections are embedded with a legal-domain model and stored in Postgres with pgvector. A question retrieves the closest sections, and the model may answer only from that text, with numbered citations. When the corpus doesn’t hold the answer, it says so.

        evaluation
        10 of 10labeled questions hit in the top five
        pass mark
        80%, scored as hit@5
        embeddings
        voyage-law-2, 1,024 dimensions
        store
        Postgres with pgvector, HNSW cosine index
        generation
        Claude, from the retrieved text only
        answers
        cited, or refused
        kind
        public code, 2026

        how a question is answered

        1. Statute corpusGeorgia code titles and closing guidance, split by section
        2. Legal-domain embeddingsvoyage-law-2 turns each section into a vector
        3. Vector storePostgres with pgvector, HNSW cosine index
        4. Closest sections retrievedtop matches by cosine similarity to the question
        5. Cited answer, or a refusal

        The rule the model runs under

        answer ONLY from retrieved sections, bracket citations, refuse if absent

        Every layer ships with offline tests and a scored evaluation.

        #method

        How I work

        • Tests and a scored evaluation ship with every build.

          I set the pass mark before I see the results, and I don’t move it afterward.

        • A person approves every change.

          An agent can write the code. It can’t decide that the code ships.

        • Every action leaves a record.

          Tickets, builds and approvals go to an audit trail, so I can show what happened and who allowed it.

        • Anything that runs unattended gets broken on purpose.

          I run kill tests before I trust a system to work while I’m away.

        #background

        Background

        I started as a bank teller in 2013 and worked my way up through relationship banking and private banking at two large U.S. banks. Since 2024 I’ve been a vice president and fiduciary officer, where colleagues bring me their hardest retirement-account questions. In January 2026 I founded Vectis Consulting to build AI workflows for small firms in regulated industries.

        Roles by year
        2026–nowFounder and CEO, Vectis Consulting
        2024–nowVice President, Specialized Fiduciary Officer
        2022–2024Senior Premier Banker
        2019–2022Private Banking Relationship Manager II
        2017–2019Relationship Banker
        2013–2016Teller
        code
        Python, FastAPI, TypeScript, React, Next.js, SQL
        data
        Postgres on Supabase, pgvector, Vercel, n8n, Git and GitHub
        AI
        Claude API, OpenRouter, MCP servers, retrieval pipelines, evaluation design, AI coding agents, local models
        education
        B.S. in Political Science, Georgia State University
        exams passed
        SIE, Series 7, Series 66
        speaks
        English, Romanian, Spanish