Tenure
AI Knowledge Engine
- year
- 2026
- role
- solo
- duration
- 9+ months
- status
- in-progress
- stack
- 15 tech
- modules
- 4
A vague mandate — "make our internal corpus feel like sitting with an expert" — turned into a working AI knowledge engine in production, built end-to-end by a single engineer. Hand the system any document corpus (12+ formats, plus IMAP / Gmail / Outlook / Dropbox sync) and it serves back persona-controlled answers grounded in that material.
The architectural bet that paid off: relevance logic lives in one isolated pod, so the orchestration layer has zero knowledge of scoring. Swapping ranking strategies doesn't touch the API. The result is a ~330K-LOC codebase that has stayed shippable across nine cycles of scope expansion — with hybrid BM25 + vector + reciprocal rank fusion behind a clean external API, voice interaction (STT + TTS), pluggable personas with per-conversation locking, and audit logging on every domain event.
Engagement is private. Sole engineer. Architecture, implementation, infrastructure, ops — alone.
The pieces that make the whole.
API Orchestrator (NestJS)
The single brain. 12 modules (auth, ingestion, conversation, persona, course, email, cloud-storage, voice, audit, api-key, inspect, v1). Owns Step Functions invocation, S3 lifecycle, pod HTTP clients, and the public v1 API. Knows nothing about how relevance is scored — that's a pod's job.
Retrieval Pod (Hybrid BM25 + kNN + RRF)
OpenSearch behind a thin Python service. BM25 for lexical, kNN for semantic, Reciprocal Rank Fusion to combine them, neighbor expansion for surrounding chunks. Deterministic dedup via sha256(courseId:content). All ranking strategy lives here, so the rest of the system stays stupid about scoring.
Ingestion Pipeline (Step Functions + Lambdas)
5 esbuild-bundled task handlers chained by AWS Step Functions: parse → chunk → embed → index → finalize. Format validation at the boundary, before anything hits the database. PyMuPDF + Tesseract OCR for the formats most pipelines pretend don't exist (scanned PDFs, mixed layouts, multi-column docs).
Email & Cloud Ingestion
IMAP, Gmail, Outlook, Dropbox. Recursive folder walking, MIME negotiation with HTML-first preference, dedup, full-sweep deletion that keeps S3 + OpenSearch + Postgres in sync. The unglamorous half of any real RAG product, built once, correctly.