L0 · Input L1 · Conv2D L2 · Graph L3 · Attention L4 · Agents L5 · WebZero L6 · Output
x = [ curiosity, data, grit ] · forward( )

ANIRUDDHA MANDAL

The Data MacGyver. I turn ML research into production systems — fine-tuned LLMs, knowledge graphs, multi-agent runtimes. Now engineering agentic infrastructure at WebZero and founding SocialTokn. Scroll to run the forward pass.

7AI companies
10+agents in production
#15EQBench3 · world
179sites benchmarked
scroll = forward pass
input · 10×10 kernel 3×3 · stride 1 feature map maxpool engineer ·0.97 artist ·0.02 other ·0.01
Layer 01 · Conv2D — 2019 → 2023

Learning to see patterns.

B.Tech in CS with Data Analytics at The Neotia University — 9.0 GPA (gold medallist) and 3rd place in Amazon's ML Challenge 2021. First production reps at Echo (later acquired by Spreetail): marketplace scrapers and automated review pipelines in Databricks and PySpark. Volunteered with Omdena, building car-damage segmentation models so MyCover.ai could insure vehicles more effectively. Like a convolution, one small kernel — curiosity — swept over everything.

PythonTensorFlowDatabricksDjangoSQL
▸ GPA 9.0/10 · Amazon ML Challenge #3
query generate h⁽ᵏ⁺¹⁾ = φ( h⁽ᵏ⁾, ⊕ msg(neighbours) )
Layer 02 · Message passing — 2023 → 2024

Everything is connected.

Three companies, one obsession: knowledge graphs. Parsed legal contracts into Neo4j for causal reasoning at LegalGraph.ai, built one of India's first disease–symptom graphs for prediction of differential diagnosis, with data collected from 20+ hospitals at Healtether Health while leading five engineers, then architected real-time GraphRAG with contextual updates at Maura.ai.

Neo4jRAGGraphRAGFastAPIGPT-4o
▸ 20+ hospitals · 1 novel clinical graph · 3 RAG systems
Attention(Q,K,V) = softmax(QKᵀ/√dₖ)·V attention weights listening…
Layer 03 · Self-attention — 2025

Teaching machines to listen.

Founding AI Lead at Vibe AI. Built a 65,356-record synthetic empathy dataset with three different LLMs and trained the world's #15 conversational model on EQBench3 using RLVER for intent based routing, then quantized it until it was fast enough to feel present (TTFT ~200ms). Attention, it turns out, is also all humans need.

Fine-tuningEQBench3QuantizationSynthetic datavLLM
▸ #15 on EQBench3 · 65,356 records · team of 3
CEREBRO multi-agent runtime msg GPT-5-mini 2.1 s LFM2.5-1.2B · <200 ms ~15× cheaper
Layer 04 · Multi-agent runtime — Jan → Apr 2026

One model became many minds.

At Sol Foundry, architected Cerebro: a runtime of 10+ specialized LLM agents classifying, extracting and enriching every message on the Sol platform, guarded by 27 automated graders with CI quality gates. Distilled GPT-5-mini into a self-hosted 1.2B model with QLoRA: 2.1 s → under 200 ms, ~15× cheaper. Turning conversations into callouts, in near-realtime.

LangChainCerebrasQLoRALangSmithDistilBERT
▸ 10+ agents · 27 graders · 87.2% across 6 languages
any website captcha wall 0 w0 score
Layer 05 · Agentic infra — Apr 2026 → now

Rebuilding the web for its next users.

The internet's newest users aren't human. At WebZero, I built and shipped w0 Score, the Webzero Agent Readiness Index: can an AI agent discover, understand and safely finish real journeys — sign-up, checkout, support — on your site? A hybrid score across Discovery, Action Surface and Governance, blended with live agent outcomes from AgentTrek, proven on a 179-site broker benchmark. Measuring the web-agent traffic, and building the web-infra for them.

MCPA2AOpenAPIAWS ECSLLM judge
▸ 179 sites · 6.5× faster evals · 153-value verdict lattice
score your site → webzero.ai
shared state human gate publish while true: research → plan → draft → measure
Layer 06 · Output — founding

An agency that never sleeps.

Founder of SocialTokn: an AI creator agency. A creator sets goals; a flat team of persistent specialist agents researches, plans, drafts and measures, continuously, from shared state — escalating only the decisions that matter to a human. The forward pass continues.

Agent runtimeA2A hand-offsDurable taskspgvector