$ ls -la ./blog/_
The four engineering decisions that separate a working voice agent from a demo that falls apart on the first real call. Vapi + GPT-4o + Supabase, shipped for Hair Time Salon in Franklin Park, NJ.
The same SKILL.md file works in both Claude Code and OpenAI Codex. Here's the recipe, the file layout, and the gotchas — including the 2% context cap most users hit.
Fine-tune a 1.7B model with QLoRA and Axolotl to match GPT-4 on structured scoring tasks at 900x lower cost. Full pipeline from data prep to vLLM deployment.
Design multi-objective reward functions for GRPO that combine exact match, off-by-1 tolerance, MAE, and format compliance to push scoring accuracy from 70% to 80%+.
Benchmark LLM inference with vLLM: measure TTFT, TPOT, throughput, and optimize with prefix caching. Includes A40/A100/H100 comparisons and capacity planning formulas.
Bi-encoder for fast retrieval over millions of documents, cross-encoder for precise reranking. A practical guide with training code, multi-GPU setup, and the two-stage pattern.
How recruitRyte uses Qdrant vector search, hybrid retrieval with dense and sparse vectors, and cross-encoder reranking to search 800M+ candidate profiles in under a second.
Manage Google Ads from the terminal using Claude Code and Playwright. Campaign creation, performance reporting, and conversion tracking without touching the UI.
An AI investment thesis for Indians: hedge career disruption by investing in Nvidia, MSFT, and AI ETFs. Covers RBI LRS rules, tax implications, and platform options.
India's misaligned incentives analyzed through the lens of AI alignment: reward hacking, specification gaming, and distributional shift in education, jobs, and society.