30
PAYING CUSTOMERS
~$4k
MONTHLY REVENUE
150+
DISCOVERY CALLS
~/johnson $ whoami
Systems for
machines that learn._
I'm a software engineer building the systems that keep AI fast and reliable in production: inference serving, agent orchestration, and event-driven data pipelines. Currently at Coalition, and active in GPU MODE writing and benchmarking CUDA/Triton kernels.
INFERENCE SERVINGORCHESTRATIONDATA PIPELINESCUDA / TRITON
cluster online
Where I have shipped
// 01 experience- 06/2025 – PresentSan Francisco, CASoftware EngineerCoalitionstack: Flask, Go, AWS, Temporal, LLM tooling (MCP), Datadog—Built a broker email intake pipeline as an event-driven ingestion system (inbound email → 8-class classifier → Temporal orchestration → attachment parsing → submission prefill), reclaiming ~35 hrs/day (≈4 FTE) of manual triage; made retries safe with Message-ID idempotency keys and S3 attachment offload.—Built Pitch Genius end-to-end (pipeline → LLM synthesis → rendering → rollout): a fault-tolerant engine fusing claims, coverage, and client-need data into broker-branded presentations, scaling to 3K+ decks and cutting prep 3 hrs → 2 min, with p95 held flat under 5× load via bounded-concurrency offloading and timeouts.—Developed Quotes-MCP (schema design → containerized deploy → org-wide rollout), exposing quote data as 30+ agent-consumable tools, adopted by 50+ engineers & PMs and cutting product-support tickets 35% (40 → 25/wk).—Co-led the short insurance application end-to-end (config schema → dynamic document rendering → cross-team rollout): a config-driven form engine for batch document generation, cutting applicant fields 35 → 10, processing 22M+ applications across 7 geos with documents persisted to S3, driving 25% quoting growth.
- 05/2024 – 08/2024San Francisco, CASoftware Engineering InternGenentechstack: FastAPI, PostgreSQL, LangChain, TypeScript, Milvus, DynamoDB—Built a clinical-trial concept management app (Postgres + trigger versioning → 14-tool LangChain agent), replacing 600+ Google Sheets with a system of record whose agent read and mutated protocol concept tables through Pydantic-validated tools, versioning every mutation as a revertible revision; 96% tool-selection accuracy.—Architected and deployed a multi-modal RAG pipeline over clinical-trial protocols using Milvus (HNSW indexing) and NLTK-optimized semantic chunking; tuned HNSW params against recall/latency/memory tradeoffs with a labeled query set and recall@k baselines, improving retrieval accuracy 45% and cutting protocol analysis from 40 to 14 hrs/week.
- 08/2023 – 05/2024Lexington, KYSoftware & Data Engineering Co-OpValvoline (R&D)stack: Python, Spring Boot, Airflow, PySpark, TypeScript—Built and shipped an ML pipeline predicting Remaining Useful Life for 300K+ vehicles: streamed 3M+ sensor data points from 5 sources through Kafka, batch-processed with parallel PySpark orchestrated by Airflow into S3, engineered features and trained with cross-validation to 98% accuracy, deployed for live inference in production.—Expanded the appointment-scheduling module using Spring Boot and Apache Kafka for real-time task queuing.
Systems I've taken across the finish line
// 02 projects
spanish mocktest ai
02 — Spanish MockTest AI
A speaking tutor that runs on our own servers
Questions I chased down
// 03 researchconnecting…
loading quotes…
Something to play with
// 04 interactLive GPU rental rates (ComputePrices) and U.S. industrial electricity prices (EIA): the two numbers that decide what training actually costs. Hover the tape to pause it.
The one-page version
// 05 resumeresume.pdf
$ open resume.pdf