Hard Problems,
Solid Foundations

A small (8B, 4-bit) Llama-3.1 fine-tune that writes first-person fiction from a dog's sensory point of view — scent, sound, pheromone — instead of the visually-dominant human frame that general LLMs default to. Trained on dog-POV literature, canine olfaction papers, and a dog-behaviour Q&A dataset.

1. 01Six-stage pipeline: ingest → chunk → embed (FAISS) → SFT-generate via Claude Sonnet Batches → LoRA train on MLX → generate.
2. 02Persona + contract enforced via cached system prompts; pillar chunks (style / science / behavior) retrieved per stimulus and injected as strict context.
3. 03Runs locally on an M2 Max after a one-time training spend — no per-token inference cost.
4. 04Adult and kids-v1 (ages 4–8) adapters from the same pipeline via a config swap; 11 versioned pilots with measurable kept-fraction deltas.

• Python
• MLX
• LoRA
• Claude Batches
• FAISS

A prototype Complex Event Processing engine for payment streams. Risk and compliance analysts write fraud rules in plain English — "if a user deposits $5,000 via ACH and tries to withdraw to a crypto wallet within 10 minutes, flag for manual review" — and the system compiles them to Flink/KSQL that runs on a low-latency Kafka stream. Zero-latency compliance: the people who understand the risk write the rules.

1. 01LLM is strictly a compiler: English → JSON AST → streaming SQL. The AI never sits in the critical path of money movement; Kafka + Flink do the actual eval.
2. 02Pluggable action sinks: liveness/SMS selfie check, 3DS step-up, API-tier throttling, payout freeze for manual review.
3. 03Every generated rule lands in Shadow Mode (logging only) for 1 hour to measure false-positive rate before a human promotes it to Active.
4. 04Targets <3 s rule compilation and <50 ms event processing latency; the LLM is out of the hot path by design.

• Kafka
• Flink / KSQL
• LLM compiler
• Streaming SQL
• AST