work

what i've built, where, and why.

quickcall.dev

/Founder
  • Compounding intelligence for agentic engineering teams
  • Contextual memory across every agent, tool, and session
  • Universal capture — Claude Code, Cursor, Codex, Gemini CLI
  • Antipattern detection, per-dev reports, auto rules generation

Yral

/AI Consultant
  • RAG content moderation — Phi-3.5 4B, 86.8% accuracy (+26%)
  • LLM inference optimization — 70% KV cache hit rate, 150 tok/sec
  • Real-time recommendation system — online learning, million-scale
  • Candidate generation pipeline — Airflow, BigQuery, Redis
  • MuseTalk video generation — image-to-video, video-to-video

People+AI

/AI Consultant
  • Evaluated 15+ Indic LLM benchmarks across 22 languages
  • Designed India's first standardized LLM leaderboard framework
  • Authored 10 trillion token collection roadmap for Indian languages

smallstep.ai

/Founder
  • Misal — India's first competitive Marathi LLM, 7B/1B params
  • Custom tokenizer — 15K tokens, fixed 3-5x Llama inefficiency
  • Beat GPT-3.5 on reading comprehension
  • Open-sourced models, tokenizer, configs, eval framework

Covered by YourStory, Moneycontrol, and ARB Podcast. Models on Hugging Face with open weights.

Medpiper

/AI Consultant
  • Medical document digitization platform
  • Auto-annotation system for labelling + training
  • 10x processing speedup (10 min → 1 min)

Tekion

/Data Scientist
  • Document AI — Mask-RCNN table extraction
  • Service recommendation engine — per-vehicle, seasonal patterns

Pratilipi

/Data Scientist
  • Autoencoder collaborative filtering — user-item embeddings
  • Content hooks — 2x reads growth post-completion
  • Author follow recs — 20% increase from profile page
  • Category personalization — top line + monetization impact