Independent AI Engineer | 01/2024 - Present
HKSoka · LLM/RAG Systems (hksoka.com/en)
LLM chat platform with production-grade memory architecture and monetisation stack — built and deployed end-to-end. Document-scale RAG: proposition-chunked ingestion via AWS Lambda — segments large seed documents, extracts atomic propositions via Haiku, embeds via Gemini into pgvector; no document size cap.
- Multi-layer RAG memory system: user-seeded long-term context + auto-learned facts, with separate retrieval quotas to prevent relevance dilution between memory types; nightly critical seed extraction — distils highest-stakes facts from learned memory into a persistent seed injected directly into system prompt, ensuring safety-critical context survives context window pressure; diagnosed and resolved cross-lingual embedding mismatch — bilingual embedding design combines source proposition with English translation in a single vector for Cantonese-English retrieval
- Embedding pipeline: AWS Lambda (Node.js) orchestrates async chunking and embedding on document save; decoupled from Vercel serverless to avoid timeout constraints
- Token cost management: token-based context window management replacing message-count truncation; conversation summarisation pipeline with incremental merge — preserving full session history across re-summarisation cycles; prompt caching on stable system prefix and conversation history, separate from dynamic RAG retrieval — 54% per-turn cost reduction on short conversations, 80–90% on long sessions.
- File understanding: PDF/image upload via base64 → Claude document API, enabling document Q&A without storage
- Web search integration: multi-turn tool-use loop with token safeguards
- Full monetisation stack: Stripe subscriptions, token-based credit system, referral incentives
- Agentic orchestration built on native Claude API — direct tool-use loop, context management, and retrieval pipeline; lower latency, full debuggability
- Stack: Claude Sonnet 4.6, Neon PostgreSQL, Vercel serverless, React/TypeScript
Ashna — Vector Similarity Matching Platform | 2026
- Embedding-matching platform: LLM-generated user portraits (Grok); Gemini embeddings + pgvector cosine similarity ranking; UMAP 2D member visualisation; Stripe payments; Vercel serverless. Shipped end-to-end.
ML Trading Pipeline · Full Stack Ownership
- Built and deployed production ML pipeline with automated signal generation, multi-seed validation, and time-series cross-validation — serving live automated decisions on a continuous basis
- Owned full pipeline from data ingestion to model deployment and AWS infrastructure
- Validated and rejected multiple theoretical optimizations based on empirical live system data
- Designed statistical validation framework to distinguish genuine signal from noise before production deployment — replacing standard backtesting with multi-seed stress testing and 3D parameter stability analysis
- Developed deployment gate system with multiple quality filters including out-of-sample median validation and minimum data coverage requirements
- Engineered dynamic exposure management system with automated adjustment and real-time monitoring
- Built production infrastructure on AWS: automated scheduling via EventBridge and real-time Telegram alerting for live system monitoring
- Research documentation available on request
