Shrey Kadia
Senior AI Engineer — LLM Systems · Agentic AI · Real-Time Voice AI · Evaluation
I build production AI systems and dig into how they work underneath the abstractions.
About
I'm a Senior AI Engineer based in Ahmedabad, India. I started out in classic machine learning, building loan-risk models, and moved into LLM systems once they became real enough to ship.
For the past year I've been building the AI layer of Mitoo, a multi-tenant business platform, on my own: an assistant that does real work across six apps, an AI agent that phones sales leads, and an AI voice interviewer. Most of that work isn't the model. It's making something probabilistic behave inside software that has permissions, tenants, money and real users.
I like going down the stack: understanding what happens underneath the APIs, not just calling them.
Selected work
- 01A permission-aware multi-agent assistant for a multi-tenant SaaSRecruiter tasks: ~1.5 min → ~15 sAn AI assistant that does real work across six business apps, acting as the signed-in user and never doing anything their role doesn't allow.NestJSTypeScriptAWS Bedrock (Claude)PostgreSQLpgvector
- 02A self-hosted AI phone agentEvery new lead called within 15 minAn AI agent that phones every new sales lead within minutes, holds a natural conversation over the real phone network, qualifies the lead and files the call in the CRM, on self-hosted telephony with no third-party voice platform.FreeSWITCH / SIPNestJSAWS Bedrock (Claude)Real-time STT / TTSAWS
Technical lab
- 01Autograd from scratch, and how small ε should befloat32 gradient checks miss bugs of 1e-4 and smallerA reverse-mode autograd engine in about 200 lines of NumPy, and an experiment on gradient checking. In float32, a perfectly correct gradient can come out 25% wrong.PythonNumPy
- 02Attention from scratch, and why it divides by √dWithout √d, the largest learning rate that trains drops 10x as heads growAttention and a full transformer block on top of my NumPy autograd engine, and an experiment on the √d in attention. My first explanation for what it does turned out to be wrong.PythonNumPy
- 03A tiny GPT from scratch, and why it needs warmupWithout warmup, two Adam steps make the model ignore the textA 0.82M-parameter GPT trained on Sherlock Holmes with my NumPy engine, and an experiment on learning-rate warmup. The loss spikes it's meant to prevent never showed up. The model got stuck instead, and I traced why.PythonNumPy
- 04A KV cache from scratch, and what it costsAt 2,048 tokens: ~2–3 ms per token instead of ~1 sA KV cache for my NumPy GPT, and an experiment on speed and memory as the context grows. The memory only becomes the problem once you serve many sequences at once.PythonNumPy
Experience
Sept 2025 — Present
Senior AI Engineer · Mehta Softech
Building the AI layer of Mitoo, a multi-tenant business SaaS, end to end and on my own. A multi-agent assistant that works across 6 apps with ~640 tools and cut common recruiter tasks from ~1.5 min to ~15 s. A self-hosted AI phone agent that calls every new lead within 15 minutes. An AI voice interviewer that has run 50+ screening interviews in 5 languages. Plus a serverless interview-evaluation pipeline and a QLoRA fine-tune of Qwen2.5-3B with its own eval harness.
Agent platform →AI phone agent →
AWS BedrockNestJSPostgreSQLpgvectorFreeSWITCHLiveKitStep FunctionsQLoRA
Apr 2023 — Sept 2025
Data Scientist · Grownited
Built and deployed loan-risk scoring models end to end, from feature engineering through training, experiment tracking and production APIs. Improved accuracy by 15%, cut false positives by 18% and brought loan-approval turnaround from 48 to 36 hours.
LightGBMMLflowFastAPIAWS EC2SQL
Education
B.E. Computer Engineering
Gujarat Technological University · 8.6 CGPA · 2023
Certification
Oracle Cloud Infrastructure Generative AI Professional · 2024
Contact
Want to talk about LLM systems, agents, evaluation or voice AI? Email is the fastest way to reach me.