Sagar Sarkale

ai founder · consultant · engineer

I've trained LLMs from scratch, built agentic systems that ship, and optimized inference pipelines that run in production.

Currently building QuickCall.

Previously built India's first competitive Marathi LLM.
covered by YourStory, Moneycontrol & ARB Podcast

book a call →

What I Built

quickcall.dev

current

Founder

contextual memory for agentic engineering teams — captures what ships, loads what matters. used across Claude Code, Cursor, Codex, Gemini CLI.

quickcall.dev

misal — marathi llm

shipped 2023 – 2024

Founder, smallstep.ai

india's first competitive marathi llm. 7B & 1B params trained on 2B+ tokens. custom tokenizer that fixed llama's 3-5x inefficiency. beat gpt-3.5 on reading comprehension. open-sourced everything.

smallstep.aihuggingfacedeep dive

Consulting

Yral2025
RAG moderation +26% accuracy, 150 tok/sec inference, real-time recs at scale2025
People+AI2024
designed india's first standardized LLM leaderboard across 22 languages2024
Medpiper2024
10x medical document processing (10 min → 1 min)2024
Tekion (2022–23)Pratilipi (2020–22)·full history

Writing

#Foundations3 posts

RNNs → LSTMs → Transformers. the learning-from-scratch track.

#Transformers6 posts

attention, position encoding, KV caching, BLT, multimodal.

#MCP3 posts

model context protocol — primitives to elicitation. 3-part series.

#Deep Dives9 posts

long-form breakdowns — vectorDB, LoRA, BLT, KV caching.

all posts →