Jigon Y.

I build data and AI automation you can check, not just trust. How I work with clients. Fixed price by scope, agreed before anything starts and split into milestones with one named deliverable each, so the cost is settled before the work is done rather than counted after it. Fully async, in writing - no live calls. Handover and tests ship with the work, so nothing depends on me still being around. What I do: • Data for automated and physical systems - sensor logs, run telemetry and event streams treated like any other production data: scheduled ingestion, schema validation, drift detection, and failure logging that surfaces a bad run instead of averaging it away. If a machine produced the number, your pipeline should be able to tell you when that number stopped meaning what it used to • Web scraping and data extraction - messy sources into clean, validated Excel/CSV/JSON, with the rows that need a human flagged instead of silently dropped • Korean-language data work - native Korean, handled as engineering rather than translation: OCR quality control on Korean documents (spacing around 조사 and 의존명사, character confusion, reading order in tables and vertical text), 자모 NFC/NFD normalization before it quietly breaks your deduplication and accuracy metrics, and structured extraction from Korean sites, forms and records • Python ETL and pipelines (Postgres/Supabase) - scheduled runs, idempotent loads, and failure logging that actually surfaces failures instead of hiding them • Document and LLM work - invoice and statement extraction with reconciliation checks, RAG question answering that cites its source and refuses when it can't, plus guardrails and evaluation harnesses that turn "looks fine" into a number What you get: every deliverable ships with the evidence behind it - what was checked, what failed, and what a human should review. If something can't be done honestly at the scope or price you have in mind, I will say so before we start rather than quietly cut corners. Ongoing work: most of what I build runs on a schedule, which means it needs someone watching it after launch. I take retainers for monitoring, drift checks and scheduled refreshes on pipelines I built or inherited. Clients usually start with one bounded piece and continue from there - my current client is on his third contract with me. How I work: fully async, in writing. I use modern AI tools to move faster on drafts and boilerplate, and I personally review, test, and stand behind every deliverable. You get speed and a human who owns the result. Recent work: an NFL game-prediction model moved from spreadsheets to a live site, a read-only BigQuery warehouse audit, and duplicate-side-effect-desk - an evaluation environment on Prime Intellect's Environments Hub that measures how often AI agents pay the same refund twice (14 vs 3 vs 3 duplicates across three models, reproducible from a fresh clone). Writing: I publish what my own checks got wrong - recent dev.to posts on the reward hacks I found in my own benchmark and on a counting bug in my own cost analysis. Proof before you hire: my case studies are public, and each one ships with its full source code and the measured numbers attached - what was checked, what it caught, what it missed. Several were built to fail on purpose: I plant known defects in the input, then publish whether the checks caught all of them and how many false alarms they raised, because a checker that never cries wolf and a checker that cries constantly are both useless. Open any project in the portfolio below and the code and the results are linked from it. Most recent, all public and reproducible: warehouse-quality-gate is a dbt contract on DuckDB - the sabotaged batch loads with zero errors and reports $4,905,051 of revenue instead of $395,751, and the contract fails 12 of 15 tests so the mart is never built. dag-guard reviews Airflow DAGs statically with ast, no Airflow install and no imports - 12 of 12 planted defects caught, including a catchup setting that queues 90,816 backfill runs the moment it deploys. metrics-contract prices definition drift between dashboards instead of just naming it - net revenue 697,691 by the contract, 755,388 on the dashboard, +8.3%. fhir-quality-gate checks FHIR R4 bundles that are already structurally valid - E119 written instead of E11.9 silently drops 3 patients out of a 46-patient quality measure, and nothing errors.