Open to Remote · EU & US Companies
AI ENGINEER · COMPUTER VISION + LLM SYSTEMS

Lakshya
Kaushik

I build production AI systems that companies pay to deploy — computer vision pipelines, multi-agent LLM workflows, and document intelligence. Shipped across three continents.

India · IST (UTC+5:30)
Available 4-5h overlap with CET
German exchange · THI Ingolstadt
6
Production AI
systems shipped
3
International
client markets
0
Hands-on AI
engineering
Egypt · Healthcare AI
Philippines · Computer Vision
United States · Infrastructure AI
Multi-market · Document AI

// selected work

Production Systems
Built & Deployed

01 / Healthcare · Egypt Market

Arabic Insurance Claims Automation

↗ Deployed for Egyptian healthcare insurance provider

End-to-end AI pipeline that reads Arabic insurance cards, identity documents, and hospital bills — then automatically matches patient data, verifies coverage, and submits structured claims. Replaced a fully manual claims processing workflow.

~90%
Manual claims
processing eliminated
Key Challenges Solved
  • Arabic OCR with mixed numeral systems (Arabic-Indic digits)
  • Female patient name matching — cards show husband's name, not patient's
  • Employer-issued vs personal insurance card reconciliation
  • Cross-document validation (insurance card + national ID + hospital bill)
  • Playwright form automation → migrated to structured JSON API submission
Tech Stack
Azure OCR OpenAI GPT-4o Python FastAPI Playwright JSON APIs Arabic NLP
Insurance Card
Azure OCR
GPT-4o Matching
Validation
Auto Submit
Claim Filed
02 / Infrastructure · United States Market

Traffic Management Automation Platform

↗ Deployed for US-based traffic engineering company

Full automation of a previously manual traffic planning workflow — from GPS coordinates to a signed-off AutoCAD traffic management plan. The system generates MUTCD-compliant road work plans with cone counts, signage placement, and staffing requirements based on state-specific guidelines.

End-to-end
Coordinates →
signed traffic plan
What I Built
  • Satellite imagery pipeline: coordinates → Google Maps API → tiled 7K resolution road image
  • Fine-tuned CV model detecting lanes, turn signals, bike tracks, speed markings, dividers
  • SAM 3 (Meta) integration for prompt-based road feature segmentation
  • Automated DXF/DWG file generation with road signs and curve geometry
  • Multi-agent LLM (Gemini + Google Grounding Search) for all 50 US state regulations
  • Final plan output: cone count, signage positions, staff requirements per state law
Tech Stack
Python Google Maps API SAM 3 (Meta) Custom CV Model Gemini Google Grounding AutoCAD DXF/DWG FastAPI Multi-Agent LLM
GPS Coordinates
Satellite Tiles → 7K Image
CV Model + SAM 3
DWG Generation
LLM Agent
Traffic Plan
My Role

What I owned on this project

Led  — Base Map Generation module
Built  — State regulation agent bots
Team  — Model fine-tuning & planning engine

I led the base map module — the pipeline that turns raw GPS coordinates into a clean, scaled, machine-readable road map — and independently built the agent system that pulls state-specific traffic regulations. The fine-tuned detection model and the downstream planning engine were owned by other engineers on the team; I worked closely with both, including helping prepare annotation guidance for the model training.

The Problem

Traffic plans were drawn by hand, one road at a time

A US traffic engineering company produced temporary traffic control plans manually — an engineer studied a road, traced its geometry into AutoCAD, looked up the relevant state highway regulations, and worked out how many cones, signs and staff a job needed. Each plan took hours of skilled work and the process didn't scale with their order volume.

The goal: take a location, automatically reconstruct the road, and generate a regulation-compliant draft plan an engineer only has to review — not build from scratch.

Architecture

Coordinates → reviewed traffic plan

GPS Coords
lat / long input
Satellite Tiling
Maps + Roads API
stitched to 7K
Detection Model
fine-tuned CV
+ SAM 3
Vector Build
lanes → DXF/DWG
signs placed
Reg Agent
Gemini + grounding
per-state rules
Plan Engine
cones · signs · staff
Engineer Review
human checkpoint
Modules I led / built
Team-owned modules

Simplified intentionally — this shows the system's logic, not the client's internal infrastructure.

Engineering Challenges

The hard parts of the base map module

Real-world scale accuracy. A satellite snapshot has no inherent sense of metres. I pulled tiles at a fixed zoom and known ground-sampling distance, then stitched them so the final ~7K image had a consistent, calculable real-world scale — without that, every downstream measurement (lane widths, curve radii) would be wrong.

Resolution vs coverage. One API call couldn't give both a wide enough road span and the detail the model needed. I built tiling logic that captured overlapping high-zoom sections and merged them into a single seamless high-resolution canvas.

Vector fidelity. Model output (coloured masks for lanes, dashes, turn arrows, bike lanes, dividers) had to become clean AutoCAD geometry. I handled the conversion into DXF/DWG so curves and lane markings landed as proper engineering vectors, not pixel blobs.

50 states, 50 rulebooks. Traffic control rules differ by state and road type. Rather than hard-code them, I built an agent grounded to official state sources that returns a fixed-format answer set (work hours, permits, emergency contacts, staffing) for any given address.

Engineering Tradeoffs

Decisions and why I made them

Why a grounded LLM agent for regulations, not a hard-coded rules database?
A static database of 50 states' rules would be accurate the day it shipped and stale soon after — and a nightmare to maintain. Grounding the agent to official state sources meant rules stayed current and the system extended to new states without a code change. The tradeoff: I had to constrain it to a fixed source list and fixed output format to keep answers reliable and prevent hallucination.
Why SAM 3 alongside a fine-tuned model instead of one model?
The fine-tuned model was strong on the specific markings it was trained on; SAM 3 added prompt-based segmentation for features that were rarer or harder to pre-label. Using both gave broader coverage than forcing a single model to do everything — at the cost of a heavier pipeline, which was acceptable because this runs as a batch job, not real-time.
Why stitch tiles to 7K instead of using one lower-res image?
Lane markings and dashed lines are tiny relative to a road. A single low-res pull lost them entirely. Higher resolution dramatically improved detection recall; the cost was more API calls and stitching complexity, which I optimised with controlled overlap.
Why keep a human engineer in the loop at all?
The system drafts; it doesn't sign off. Traffic plans carry legal and safety weight, so a mandatory engineer review checkpoint was a deliberate design choice — automation for the 90% of effort, human judgement for the accountability.
Lessons Learned

What I'd carry into the next system

01

Grounding beats hard-coding for anything that changes over time. The maintenance saving compounds.

02

For CV, input resolution often matters more than model choice. Fixing the image pipeline beat tuning the model.

03

Constraining an LLM to fixed sources + fixed output format is what makes agentic output trustworthy in production.

04

Design the human checkpoint deliberately. Knowing where not to automate is part of the engineering.

03 / Legal & Enterprise · Document Intelligence

AI-Powered Contract Differentiator

↗ Built for enterprise contract management workflows

Side-by-side PDF comparison tool that uses AI to classify and highlight every difference between two contract versions — distinguishing structural changes, clause modifications, and minor corrections. Gives legal and procurement teams instant clarity on what changed between drafts.

3-tier
Change classification
system
Change Classification
  • Structural changes — clauses added, removed, or reordered
  • Semantic changes — meaning altered in existing sentences
  • Surface corrections — spelling, punctuation, formatting only
  • Page-level diff view with color-coded highlighting per type
Tech Stack
Gemini File API Python PDF Processing FastAPI React Diff Algorithms
04 / Construction · Document Intelligence

Construction Document Analysis & Roof Estimation

↗ Deployed for US construction estimation company

Multimodal AI pipeline that processes construction blueprints, PDFs and site images to extract structured data — including roof geometry, canopy dimensions, and material specifications. Calculates actual surface areas using real-world formulas to generate accurate cost estimations.

Auto
Surface area
from blueprints
Key Capabilities
  • Multimodal ingestion — PDFs, images, and mixed construction documents
  • Roof geometry extraction and real-world surface area calculation
  • Structured output in client-defined estimation format
  • Canopy, ridge, and slope data parsing from technical drawings
Tech Stack
Gemini File API Multimodal LLM Python FastAPI PDF Parsing Geometry Algorithms
05 / Surveillance · Philippines Market

CCTV Computer Vision System

↗ Deployed for Philippines-based security operator

Real-time CCTV intelligence pipeline running on edge infrastructure — loitering detection, trespassing alerts, and continuous vehicle counting. Built for 24/7 operation across multiple camera feeds.

Live
Real-time
edge inference
Detection Capabilities
  • Loitering detection with configurable dwell-time thresholds
  • Restricted area trespassing alerts
  • Continuous vehicle counting across multiple camera feeds
  • Real-time alert pipeline for operator dashboards
Tech Stack
Python YOLO OpenCV Object Tracking RTSP Streams FastAPI

// capabilities

Technical Stack

AI & ML — Core Strength
LLMs RAG Multi-Agent Systems Computer Vision OCR Pipelines Fine-tuning SAM 3 YOLO Prompt Engineering
Evaluation & Production
LLM-as-Judge Eval Harnesses Observability Cost Optimization Model Routing Structured Outputs Guardrails MCP Integration
Models & APIs
OpenAI GPT-4o / 5 Gemini Claude Azure OCR Google Grounding Gemini File API Google Maps API
Backend & Infra
Python FastAPI Docker GCP AWS Azure PostgreSQL MongoDB Supabase
Frontend — Working Knowledge
React Next.js HTML / CSS Node.js Tailwind CSS

// timeline

Experience

Jul 2025 — Present

Pune, India
AI Developer
Geeky Bee AI Private Limited
  • Built Arabic OCR insurance automation pipeline — reduced client manual processing by ~90%
  • Led satellite imagery → AutoCAD pipeline for US traffic management platform using SAM 3 and fine-tuned CV models
  • Built multi-agent LLM system with Google Grounding Search covering all 50 US state traffic regulations
  • Developed AI-powered PDF contract differentiator classifying structural, semantic, and surface-level changes
  • Built construction document analysis pipeline using Gemini File API for roof estimation and takeoff analysis
Jan 2025 — Jun 2025

Pune, India
AI Development Intern
Geeky Bee AI Private Limited
  • Designed RAG-based QA system reducing query resolution time by 30%
  • Built CCTV computer vision pipeline for loitering detection, trespassing alerts, and vehicle counting — deployed in Philippines
  • Automated data pipelines reducing manual workflows by 70%
Jul 2024 — Present

Pune, India
Research Intern
SCAAI — Symbiosis Centre for Applied AI
  • Applied AI research in computer vision and multimodal systems
  • AgNOR detection in medical imaging using OpenCV and deep learning

// academic

Education

🇮🇳 India
Symbiosis Institute of Technology
B.Tech in Artificial Intelligence
Sep 2021 — May 2025 · Pune
🇩🇪 Germany · Exchange
Technische Hochschule Ingolstadt
B.Tech AI — Image Recognition & Applied AI · Grade 2.2
Mar 2024 — Sep 2024 · Ingolstadt

Let's Build
Something Real

Open to remote AI engineering roles with EU and US companies. Comfortable with CET / EST timezones. Available for full-time positions and contract work.