// projects, ranked from "actually hard" to "please don't ask about the deadline"

► ★ FINAL BOSS: The Privacy Layer for Public LLMs

PRIVACY LAYER FOR PUBLIC LLMs ON ENTERPRISE CODE

Self-shipped · Python · MCP · mitmproxy · SQLite

A two-layer privacy framework that lets enterprises use Claude / GPT / Gemini on proprietary source code. The developer sees real code, the cloud sees pseudonyms, and an independent guardrail kills anything that slips through.

Layer 1 — Masking MCP Server. Sandboxed file API exposed to Claude Desktop. Replaces internal names with consistent tokens like [APP_001] so the model can still reason about the code, with a reversible SQLite vault for local unmask.

Layer 2 — Guardrail mitmproxy addon. Independently scans every outbound request to the LLM provider. Fail-closed. Assumes Layer 1 might be broken.

74JAVA FILES TESTED

4,062SAFE REPLACEMENTS

~20 msMASK + UNMASK

0LEAKS / 74 FILES

Framed against real industry incidents — the 2023 Samsung source-code leak into ChatGPT, the corporate ban wave (JPMorgan, Goldman, Apple, Verizon), and the Cyberhaven finding that ~11% of pasted ChatGPT content is sensitive.

► Quest 1: The Tosca Dragon

TRICENTIS TOSCA CONNECTOR · FY23-24

Flagship delivery · Customer: Roche · End-to-end ownership

Built a new connector for Tricentis Tosca on the OIM platform. The source system was so unconventional that standard modeling didn't fit — modeled test entities, test steps, workspace lock, and cloned workspaces with a custom approach. Owned dev, design, modeling, customer demos, and customer interactions. Manager flagged this as the standout delivery of the year.

► Quest 2: The Mastercard Sync Speedrun

MASTERCARD SYNC PERF · ~6× FASTER

Bug 253757 · Triage rotation win

Sync at Mastercard was bottlenecked on DB query frequency. Optimized the hot path during triage — ~6× performance improvement on the customer sync. Triage as perf engineering, not just firefighting.

► Quest 3: The 1,200-Call Problem

RALLY API CACHING · 1,200 → 1 CALL

Enhancement 199414 · Mastercard

OIM was making ~1,200 Rally API calls per cycle just for projects & type definitions. Implemented caching on both fetches and dropped it to 1 call per cycle — massively reduced load on the customer's Rally instance and sped up syncs.

► Side Quests Shipped

Intuitive Relationship Configuration in OIM (Feature 236592)
Archived Entity Link Support — Core + Jira (Feature 206780)
Enhanced user mapping in OIM4ADO (Enhancement 235197)
SDK dynamicDataRetrieval support (Feature 216937)
Time-impersonation for ADO / TFS (Bug 216902)
Rank & hierarchy for all entities in Jama (Enhancement 232713)

PAUSED

STAGE CLEAR!