Blog
Build logs, written from the work itself.
Each post comes from the actual commits and decisions of a project, not retrospective polish.
Up-Candel
A paper-trading platform with a market you can replay and rewind, and the ledger rules that time travel forced on it.
Read the postweekly-loop
A Claude Code plugin that runs a project plan one week per tick, and only believes exit codes and greps, never the agent's own report.
Read the postAnswerWala
Voice question answering in Hindi and English over 950K passages, held under 200 ms at P90 from transcript to answer.
Read the postOverdue
An app that writes the admin messages you keep avoiding and drafts the follow-up on its own when nobody replies. Built with two coding agents for OpenAI Build Week.
Read the postIndicRAG Bench
Building a benchmark for Indian-language document RAG that starts from the scanned page, not from clean translated text.
Read the postU-JEPA
Continual-learning research, including the phase where the whole approach failed and I rebuilt it.
Read the postVL-JEPA
A vision-language model trained from scratch on free Kaggle GPUs, and the contrastive collapse that nearly sank it.
Read the postAegis
Cryptographic provenance and deepfake defense for sports media, deployed on Google Cloud.
Read the postBOL BHARAT
AI-powered civic issue reporting platform using Gemini Vision for issue detection.
Read the postAI Interviewer
Recruitment automation that conducts and evaluates candidate interviews with AI.
Read the postMindLens AI
Transparent mental health risk detection for social media text.
Read the post