~/projects/stochi

Stochi

Supplement tracking with interaction warnings and pharmacokinetic modeling

# Role: Solo Developer & Designer # Duration: 3 months # Year: 2025
Next.js 16 Go PostgreSQL pgvector Web Workers Transformers.js Drizzle ORM Tailwind CSS
~/rationale

Biohackers and health-conscious individuals often take multiple supplements without understanding potential interactions, timing conflicts, or dangerous dosage combinations. Existing resources require manual cross-referencing across fragmented medical databases, leaving users unaware of risks like zinc-induced copper deficiency or iron absorption competition.

The platform required modeling complex pharmacokinetic behaviors including saturable absorption (Michaelis-Menten kinetics) for supplements like Vitamin C and Magnesium. Additionally, the system needed to provide instant search without server round-trips while maintaining a comprehensive knowledge base backed by actual research.

Stochi combines a Go microservice for sub-millisecond pharmacokinetic calculations with a Next.js frontend featuring client-side semantic search using quantized LLMs in Web Workers. The RAG pipeline surfaces research-backed answers while deterministic safety guardrails enforce FDA limits on toxic supplements regardless of AI suggestions.

~/highlights

Pharmacokinetic Modeling Engine

1,000+ supplements · sub-ms calculations
01

Built a Go microservice performing sub-millisecond calculations for 1,000+ supplements. Most follow exponential decay, but supplements like Vitamin C and Magnesium saturate at high doses, requiring Michaelis-Menten equations solved analytically using the Lambert W function.

Client-Side AI Search

90% queries offline · 23MB model
02

Implemented zero-latency search using quantized LLMs (Transformers.js) running entirely in Web Workers. This eliminates server costs for 90% of queries while enabling offline PWA support with the all-MiniLM-L6-v2 model (~23MB).

RAG Research Pipeline

1536-dim embeddings · cited answers
03

Built a retrieval-augmented generation system using Llama to rewrite medical queries, search relevant research chunks from scraped Examine.com data stored in pgvector, and generate answers backed by actual citations with automatic safety guardrails.

Deterministic Safety Layer

FDA-compliant · hard safety limits
04

Enforced toxicological safety across all supplements with hard-coded FDA limits (Zinc 40mg, Iron 45mg, Vitamin D 10,000 IU) that bound all AI-generated suggestions. Soft limits trigger warnings for low-risk supplements while hard limits cannot be overridden.