Hey, I am Yash

150-trial experiment measuring what strict JSON output contracts do to multi-step reasoning. Answer-only JSON drops Sonnet 4.6 from 25/25 to 10/25 on multi-step problems; adding a single schema reasoning field recovers 24/25 while using 59% fewer visible output tokens than free text.

Pre-registered experiment on what repeated context compaction does to stored facts. Under default summarization, facts decay with a half-life of roughly one round; a one-line verbatim-preservation instruction restores 100% retention.

A controlled study of free-running agent loops versus structured graph execution on the same workloads, with the full experiment code open-sourced.

A field catalogue of the applied AI product patterns that actually made it to production in startups between 2024 and 2026, written from the builder's side of the table.

Applying the defense-in-depth security model to language-model applications: layered controls across prompts, tools, and data instead of a single guardrail.

Memory layer for AI agents. Ran a self-directed bug audit across the codebase and turned it into 10 pull requests spanning vector-store backends and LLM integrations.

High-performance vector search engine written in Rust. Actively contributing fixes upstream.

• Built cross-platform desktop app with Electron + React spanning 5 microservices (DMG in 8 weeks)
• Real-time ONVIF/RTSP video streaming at sub-100ms latency
• Deployed 4 AI/ML models for attendance and face recognition: 98%+ accuracy on 10,000+ daily requests

• Improved continuous authentication accuracy from 89% → 92% for banking apps
• Analyzed molecular olfaction with Graph Neural Networks across 15,000+ pharmaceutical compounds

Problem: 800,000 Indian clinics run on 3-minute paper consultations, doctors forget patients the moment they leave, losing ~₹50K/month in follow-ups. Doxa is an AI scribe that listens to a 15-second Hinglish voice memo after each patient and turns it into a structured record + automatic follow-up SMS. Zero workflow change, no EMR. The startup I'm building now.

Problem: SRE work is mostly pattern-matching against past incidents, but engineers spend hours doing it manually at 3am. Pilot is a fleet of AI agents that triage incidents, parse logs, gate deploys, and audit infra hygiene, Claude-driven, runs against your real stack, escalates only when humans actually need to look.

Problem: cold-email tools either spray-and-pray (low reply rates, deliverability tanked) or require 30 min of manual research per prospect (doesn't scale). Outreach combines per-prospect research, AI-personalized copy, warm-up rotation, and reply tracking, campaigns that actually feel hand-written.

Problem: chess.com is bloated and Lichess UI feels dated to me. Built a clean multiplayer chess platform with sub-50ms move propagation, Redis-backed game state, ELO matchmaking, OAuth login, and containerized deploys. Real-time chess that just feels fast.

Problem: my photo work was scattered across Instagram, Drive, and an old Lightroom catalog. shots is a self-hosted portfolio for the street, portrait, and travel work I've shot across Bangalore and India, built so I own the canvas and the metadata, not a platform.

Problem: students keep losing job opportunities to ATS filters because they don't know what a 'good' bullet looks like. finalcv is an AI resume builder with ATS-tuned templates and Gemini-powered bullet rewrites tailored to the JD. 1,000+ monthly users, sub-second suggestions.