GenAI Analyst &
AI/ML Engineer

Autonomous agent systems often run into runaway token costs, unsafe tool execution, and erratic multi-step coordination across enterprise backends.

Built a LangGraph supervisor-worker architecture with custom MCP servers, gVisor container isolation, context compaction, human approval gates, and rigorous evaluation suites.

Token usage dropped 42% per trajectory while task completion across 500+ golden scenarios surged from 71% to 94.2% at sub-3 second P95 tool latency.

Agent demos end at a pass/fail score, which says nothing about which step broke, whether a fix regressed something else, or what a new safety policy would have prevented.

Built a Python harness with typed tool contracts, sandboxed state, canary egress checks and prompt-injection scanning, plus SQLite run storage, an authenticated review console, trace comparison and offline policy replay.

Every failure becomes an evidence capsule with a fingerprint, root cause and suggested control, and a candidate agent can be compared to its baseline case by case before release.

Engagement-only rankers collapse into popularity loops, and most recommender demos stop at an offline notebook with no serving, exploration or feedback path.

Trained a content-aware two-tower model with hard negatives and FAISS retrieval, a multi-task ranker for completion, like and skip, then added creator and category diversity caps, LinUCB exploration with propensity logging, and idempotent online feedback.

On the seeded evaluation set the model reaches 1.55x the recall of a most-popular baseline while covering 85% of the catalog, and the UI exposes those metrics next to the live feed.