Principal Systems Engineer (TypeScript & PostgreSQL)
Build the AI governance control plane runtime — database schema, hybrid search, and governance middleware between evidence and model.
Stack: TypeScript, PostgreSQL (Aurora / pgvector), React, local LLM microservices (Python/C++)
Why this role exists
Kenshiki's enforcement architecture runs on PostgreSQL and TypeScript. The bounded-synthesis pipeline retrieves evidence from Kura, applies governance gates, and either emits a verified response or refuses. All of that needs infrastructure: database functions, hybrid search algorithms, deterministic middleware, and ingestion pipelines. The taxonomist defines the rules. The reliability engineer sets the tolerances. You build the engine that executes both.
What you'll do
- Own the database layer. Design and optimize the Kura schema in PostgreSQL. Write the functions that execute Reciprocal Rank Fusion, blending tsvector lexical search with pgvector semantic search at sub-second latencies.
- Build the governance middleware. Write the TypeScript orchestrator that applies SIRE inclusion/exclusion rules to retrieved evidence — Unicode-aware matching, AST parsing, boundary enforcement — before anything reaches the model.
- Build ingestion pipelines. Transform hierarchical compliance documents into atomically chunked JSON envelopes for database insertion. Orchestrate offline entity extraction and embedding generation.
- Run local model infrastructure. Deploy and maintain the containers hosting embedding models and local inference engines (vLLM with PagedAttention), ensuring reliable, high-throughput API bridging with the Aurora backend.
What we're looking for
- Systems thinking: You write TypeScript but think like an OS developer. Memory management, efficient data pipelines, strict boundaries.
- PostgreSQL depth: Not just ORMs — you understand transaction locks, row-level security, trigger functions, and vector index tuning (HNSW vs. IVFFlat).
- High-performance TypeScript: Fault-tolerant middleware, complex async I/O, no event-loop blocking.
- Structural enforcement mindset: You understand that better embeddings do not solve hard logical negations. You build the architecture that enforces truth, treating the model as an untrusted component.
Why Kenshiki?
You'll build infrastructure where structural truth overrides probabilistic guessing — alongside domain experts who know why it matters. The system you build is the enforcement architecture.