Transformco (Sears Home Services)

AI Engineer (Contract)

August 2026 – Presentcurrent

What I did

  • Working on the 1099 contractor platform, the recruiting, onboarding, compliance, and performance-audit estate for Sears Home Services' independent appliance-repair network, spread across a 13-repo GitHub organisation.
  • Rebuilt the provider-audit agent as a single Python service in one container, running in shadow mode alongside the live agent: it pulls 90-day recall statistics per provider from Snowflake (read-only guarded), assigns RED/ORANGE/YELLOW/GREEN tiers with longitudinal counters, emits nine recall and survey event types, and records the enforcement actions it would take, executing none. A 4,248-line core with 261 tests, 14 modules, and all 36 event types at parity with the operational agent.
  • Added 30-day provider-trajectory forecasting with Prophet (linear fallback) and IMPROVING/STABLE/WORSENING banding, plus a glassbox decision-tree classifier with a RandomForest challenger (scikit-learn).
  • Shipped it on Flyway schema migrations behind an ArgoCD PreSync hook, Kubernetes dev and prod overlays, non-root Docker with a healthcheck, and SonarCloud in CI, with a reconcile CLI that diffs the shadow tier board against the operational database and writes a divergence report.
  • Built the organisation's knowledge index, a hybrid-search RAG service over the whole estate (commits, pull requests, specs, task records, contacts): LlamaIndex chunking with FastEmbed 384-dim ONNX embeddings as the vector leg and LanceDB native BM25 as the keyword leg, fused by reciprocal rank fusion, deliberately LLM-free on the retrieval path so it returns ranked sources rather than generated answers. An 11,042-row index that ingests in about 50 seconds and runs fully offline after a one-time model download.
  • Added a second sweep that harvests all 1,681 repositories in the organisation into a corpus and runs a LlamaIndex and DeepSeek map-reduce over it to produce an engineering-principles report, resumable on pushed_at with blobless sparse clones above 100MB.
  • Backed both with a 13-query golden retrieval-eval suite carrying per-query expected sources and written rationale, 62 backend tests, and 4 Playwright end-to-end specs over a 22-route Next.js 16 and React 19 front end.

Tech stack

  • Python
  • FastAPI
  • LlamaIndex
  • LanceDB
  • FastEmbed
  • DeepSeek
  • Prophet
  • scikit-learn
  • pandas
  • Snowflake
  • PostgreSQL
  • Flyway
  • Docker
  • Kubernetes
  • ArgoCD
  • AWS Secrets Manager
  • pytest
  • ruff
  • SonarCloud
  • Next.js
  • React
  • TypeScript
  • Playwright