Software Engineer · Polaris Wireless
Full-time · Santa Clara, CA
- Shipped an AI data analysis platform end to end in Python (FastAPI, LangGraph, PostgreSQL, React): a streaming agent runtime with disconnection-safe resumable runs, checkpointed state, MCP tools, JWT/RBAC access controls, human-in-the-loop approval, reusable agent skills, and an nsjail sandbox for model-generated code. Agents run governed text-to-SQL across five database engines (ClickHouse, DuckDB, SQL Server, MySQL, PostgreSQL).
- Deploy and operate the platform’s self-hosted vLLM inference on GPU: own model deployment and upgrades, tune KV-cache memory, batching, and context length for throughput and latency, and debug the serving path in production.
- Built a 133-case end-to-end evaluation and regression harness that grades each agent answer in four layers — tool routing, arguments, results, and rendered charts and maps — against a ClickHouse ground-truth oracle. It drove accuracy to 97% and cut suite runtime 57% by fixing the tool-call and grounding defects it exposed.
- Cut co-presence analysis, which scores which subscribers met a target, from two minutes to under 30 seconds across billions of mobility records by redesigning the ClickHouse pipeline, and derived and documented its statistical confidence model.
- Built and own production KPI monitoring for customer deployments in two countries using Flask, Grafana, Prometheus, Loki, and Docker: 100+ KPIs across 48 dashboards, 15 alert rules, scheduled 15-minute analysis jobs, a web command portal, and automated monthly reports.
- Developed natural-language workflow agents for the AIGIS product: intent classification, fuzzy resolution of target and geofence names, natural-language date parsing, parameter validation, and generation of structured API requests.
- Technical lead for the analysis platform: set its architecture and drove delivery with an offshore engineering team. Mentored several interns and conducted technical interviews for 10+ engineering candidates across multiple teams.