Tests whether a supervised fine-tuned (SFT) LoRA is more accurate and cheaper than calling a base model with either a detailed system prompt, or a base model + RAG. Built on Fireworks AI as the first classification step of a bank support assistant, sorting each customer message into one of 77 pre-defined intents.

• Fine-tuning beat prompting by 15.8 points (92.7% vs. 76.9%), but retrieval over the full training set tied it at 94.0%, inside the ±3-point margin of a 385-message test.
• The LoRA's real edge is cost structure: about 12 times fewer tokens per request (66 vs. about 770), which only pays for its dedicated GPU above about 15.8 million requests a month.
• Every approach is scored on the same held-out messages and reported with confidence intervals, so small gaps are not presented as wins.

A serverless relay that reads the AWS China cost and usage report, rewrites every money column from yuan to US dollars while keeping the report's exact format, and copies it to a commercial-partition bucket so CloudZero ingests it as native AWS data.

• Idempotent by design: a deterministic ID built from the source version and exchange rate means a month is only reprocessed when the data or the rate changes.
• Exact decimal arithmetic for money, verified by a round-trip test that must match the original to the cent.

A grounded ElevenLabs agent that answers from a source-tagged profile, takes actions through client tools such as showing projects and opening a booking page, plus a bring-your-own-AI onboarding kit and a validator so anyone can build their own.

• Every fact carries a source tag, and a validator blocks untagged claims, phone numbers, and placeholder text before anything ships.
• Tools for actions instead of generated text: links and bookings are deterministic.
• Voice runs over WebRTC and chat over WebSocket through one agent, under a strict content security policy.

A LLM Assistant that guides a proof-of-value from success criteria to a validated business case, with state kept in shared storage so any sales engineer can pick it up. It cuts each evaluation by more than two hours.

• Synthesizes customer calls into prioritized tasks against SOW success criteria, using focused RAG, multiple MCPs, and Python validation logic.
• Automates POV stand-up and cost allocation, cutting time-to-value by ~2 hours per POV.

A routed knowledge corpus, a custom linter with 131 tests, 20 worked example patterns, and an eval harness, packaged so Claude Code, Cursor, Copilot, and Codex all behave the same way.

• Constrain and verify instead of trusting generation: the linter catches the model when it is wrong.
• Observable compliance: agents must state which files they read before generating.
• One corpus shipped in four agent formats, kept identical so behavior matches across tools.

A two-pass culling app with intent-aware grading, taste profiles, and editorial guidance, with billing and a bring-your-own-key option.

• Downscaling for the cull pass cut vision-token cost about 60 percent.
• Measured determinism instead of assuming it: 340 API calls showed 100 percent rating stability at temperature 0.
• Regression-tested scoring with eval fixtures on every prompt change.

An event-driven serverless pipeline plus a backfill tool, built for customers to deploy and run themselves.

• Selective retries with exponential backoff, and structured JSON logging for supportability.

A Chrome extension that logs each click and navigation during a demo, plus a Python pipeline that describes each step with Claude vision, fact-checks a written narration, voices it with ElevenLabs, and assembles the final video with captions.

• Claim-level verification of the narration against frames and cited sources before any audio is generated.
• A human approval gate before every paid step, and resumable runs that only redo stages whose inputs changed.

An app that runs locally, counts what a cleanup would remove before acting, and trashes mail in batches of 1,000 with undo.