Ashutosh Bansal · Agentic AI Engineer

Experience & Education

May 2026 – Aug 2026

SDE & GenAI Intern

INDmoney · Gurgaon

Engineered production GenAI extraction pipelines for NRI bank statements and ITR documents across 5 KYC document types using MarkItDown and Gemma 4B, improving field extraction accuracy by 8–10%. Led CVL KYC infrastructure revamp integrating updated CVL APIs, optimizing payload processing, and hardening against schema shifts. Benchmarked HyperVerge and 3+ vendor document extraction platforms, improving OCR accuracy by 8% via custom preprocessing.

Mar 2026 – May 2026

McKinsey Forward Program

McKinsey & Company · Global Cohort Selectee

Selected for McKinsey's competitive global leadership program for next-generation professionals. Focus areas: practical problem-solving frameworks, structured communication, and data-driven decision making under ambiguity.

2023 – Expected 2027

B.Tech · Computer Science & Engineering

Manipal Institute of Technology, Bengaluru · CGPA 7.9 / 10

Core focus: Multi-agent AI systems, distributed computing, and reinforcement learning. Junior Placement Committee (CECAS-MITB): spearheaded corporate recruitment outreach, driving a 20% increase in company participation.

Hackathons

01Adobe Hackathon 2026 — Web & Skill Audit Orchestrator2026+

Problem Statement

Design an autonomous audit orchestration engine capable of crawling, analyzing, and validating complex web environments and AI agent skills under strict rate limits, WAF constraints, and URL isolation boundaries.

My Solution — ThreeSum Audit Engine

Built an intelligent multi-step audit orchestrator combining TypeScript, Bun, and Playwright with an official Model Context Protocol (MCP) server. Features root URL isolation, seed normalization, automated WAF block detection, and rigorous skill validation routines against official specifications.

✓ Submitted · Adobe Hackathon 2026
02Hackbricks 20262026+

Problem Statement

Build a real-time fire detection and early-warning system using open civic datasets. The challenge required integrating live government data sources to produce actionable alerts with minimal latency.

My Solution

Built a fire detection and alarming tool using live San Francisco fire incident datasets via the Socrata Open Data API. The system polls real-time incident feeds, runs a classification layer to filter fire-type incidents, and triggers graded alerts based on proximity and severity. Includes a lightweight dashboard for monitoring active incidents.

· Participated · Hackbricks 2026
03AI in Mobility — Mahe Mobility Challenge2026+

Problem Statement

Design an AI system for autonomous vehicle perception — specifically drivable space detection — under a strict constraint: no pretrained model weights allowed. All models must be trained from scratch on the provided nuScenes dataset.

My Solution — FreeSeg

Built FreeSeg: a custom encoder-decoder segmentation model entirely from scratch in PyTorch. Used a CNN encoder with residual shortcuts and a U-Net-style decoder with skip connections. Compensated for no pretrained generalisation through aggressive augmentation — random crops, colour jitter, flips, Gaussian blur. Achieved ~0.68 mean IoU on drivable surface class.

🏆 Top 10 / 90 Teams · Mahe Mobility Challenge 2026
04Databricks Hackathon — HabitEcho2026+

Problem Statement

Use the Databricks lakehouse stack to build a meaningful, production-grade data product. The challenge encouraged creative application of Databricks tools including MLflow, Delta Lake, and the Medallion architecture.

My Solution — HabitEcho

Built HabitEcho, a passive phone behaviour monitoring system. Ingested raw phone usage logs through a Bronze → Silver → Gold Medallion pipeline on Databricks. Applied Granger causality analysis to surface which app behaviours predict productivity outcomes, and used Isolation Forest for anomaly detection in usage patterns. Tracked all experiments with MLflow.

· Submitted · Databricks Hackathon 2026
05Unisys UIP Y172026+

Problem Statement

Design an intelligent IT operations solution that reduces mean time to resolution (MTTR) for enterprise incidents. The solution needed to demonstrate real agentic capability — not just a chatbot wrapper around ticket lookup.

My Solution — AutoResolve

Built AutoResolve: a 4-agent LangGraph pipeline with Triage → Diagnostic → Remediation → Verification stages. All agents communicate via a typed shared state object. Groq powers fast inference. Pre-approved remediation playbooks are executed by the Remediation Agent. Human-in-the-loop gates enforce review before any severity-1 action executes.

✓ Submitted · Unisys UIP Y17
06Zomathon — PrepSense2026+

Problem Statement

Build a product that solves a real pain point for students or early-career professionals, leveraging AI. The hackathon was organised by Zomato and emphasized product thinking alongside technical execution.

My Solution — PrepSense

PrepSense was an AI-powered interview preparation platform. It generated personalised interview question sets based on job description + resume, tracked performance across mock sessions, and surfaced improvement patterns. Used NLP to identify skill gaps between the target JD and the user's current responses, then recommended targeted preparation paths.

· Submitted · Zomathon 2026
07HackNite'25 & RILO Hackathon — SDR Lead Scraping Agent2025+

Problem Statement

Sales development teams spend disproportionate hours manually searching, verifying, and enriching high-intent inbound and outbound business leads across fragmented directories.

My Solution — Multi-Agent SDR Lead Agent

Engineered an autonomous multi-agent lead discovery and qualification workflow. Orchestrates web scraping, schema normalization, email verification, and personalized outreach draft generation with strict rate limiting and deterministic error recovery.

🏆 Top 10 / 100 Teams (HackNite'25) & Top 10 / 75 Teams (RILO 2025)

Building systems
that think, decide
& act.

About me

Tech stack

Projects

1,230 contributions in the last year

Notes & Writing

Problem Statement

My Solution — ThreeSum Audit Engine

Problem Statement

My Solution

Problem Statement

My Solution — FreeSeg

Problem Statement

My Solution — HabitEcho

Problem Statement

My Solution — AutoResolve

Problem Statement

My Solution — PrepSense

Problem Statement

My Solution — Multi-Agent SDR Lead Agent

Let's work
together.

About me

Tech stack

Projects

Notes & Writing

Problem Statement

My Solution — ThreeSum Audit Engine

Problem Statement

My Solution

Problem Statement

My Solution — FreeSeg

Problem Statement

My Solution — HabitEcho

Problem Statement

My Solution — AutoResolve

Problem Statement

My Solution — PrepSense

Problem Statement

My Solution — Multi-Agent SDR Lead Agent

Let's worktogether.

Let's work
together.