Insights
Build notes on LLM systems, backend design, and applied ML — written up from things I actually shipped.
Hitting ~950ms on a $0/month Voice Agent
How I chained Twilio Media Streams, Deepgram, an LLM with function calling, and ElevenLabs over WebSockets — and the sticky provider failover and sliding-window context that killed the dead air.
Read postEnforcing Per-User Data Isolation at the Query Layer
Why route guards aren't enough. Scoping every PostgreSQL query by userId in DSA Guardian so that no query can ever run unscoped — plus encrypting session cookies at rest with NaCl secretbox.
> read_snippetA Modular Feature Framework for EEG-Based MDD Detection
Notes from IIT Indore: building time-domain, frequency-domain, and nonlinear feature extractors into one reproducible pipeline — and validating it to 92.1% accuracy with cross-validation.
Read postTime-Aware Splits: Not Leaking the Future into Your Model
Lag and rolling features make time-series models strong — and make random train/test splits silently cheat. How I structured the stock-price pipeline to evaluate honestly.
Read post