Sound on

28°36′50″N 77°12′32″EDelhi
  • AI Engineer
  • Software Development Engineer
  • Full-stack developer
  • 3D web developer

AI agents, full-stack products and 3D websites, built and shipped to production.

Open to internships, full-time roles and client projects

Book a 30-min callDownload résumé
Loss here
—
Marble
none yet

Field notes About

Based in
Delhi, India28°36′50″N 77°12′32″E
Studying
B.Tech CSE (AI/ML) at MAIT ’27CGPA9.0+
Currently
AI Engineer Intern at Hive ProRemote · Dubai
10jainlakshya@gmail.comRésuméPDF

What he ships Capabilities

5 lines · each one live

AI agentsHuman-in-the-loop agent runHive Pro · 2026
agent platforms
Production
sign-off before it acts
Human
LangGraph · RAG
Claude
TENANT A · ISOLATEDTOOLSRBAC · APPROVERINTAKErequestPLANLangGraph · ClaudeHUBSPOTSLACKGMAILPOLICY GUARDRAILACTtool callHUMAN APPROVALAUDIT LOGEmpty. The first action is on its way.
Simulated
Paused off-screen
Waiting
0
Approved
0
Blocked
0
Guards
RBAC · tenant isolation
Nothing acts until it clears policy guardrails and a person with the approver role signs off. Wired into HubSpot, Slack and Gmail.
Search · RAGHybrid retrievalScrini AI · 2025
candidate profiles
10K+
retrieval (HNSW)
< 50 ms
cross-encoder
Rerank

Top 5searching

  1. 1c-05376.90ML eng · 10 yrs · remote
  2. 2c-08040.90ML eng · 11 yrs · remote
  3. 3c-01893.89ML eng · 6 yrs · remote
  4. 4c-06267.89ML eng · 7 yrs · remote
  5. 5c-03112.88ML eng · 4 yrs · remote
Simulated

Query “LLM engineer, RAG pipelines” Filter remote

Engine
FastAPI · Qdrant
Drawn
2,400-point sample
Semantic resume search: cosine similarity and metadata filters find the candidates, then a cross-encoder reranks them for relevance.
On-device MLSpectral velocity odometrySETU · on-device
CNN-GRU model
207 KB
inference on a phone
1–5 ms
parameters
84,000+
Simulated signal

v = 2πr · f with r = 0.31 m, a tyre radius assumed for the simulation

Wheel peak
—
Speed est.
—
Reads wheel-rotation frequency out of a phone's accelerometer vibration and turns it into speed, without double-integrating acceleration.
RoboticsLiDAR sweep · occupancy gridA.T.O.M Robotics Lab · 2024
prototype
Mars rover
+ Python
ROS2
sensor fusion
LiDAR + IMU
Simulated
Mapped
—
IMU heading
—
Path planning
re-plans as it maps
Autonomous GPS navigation with path planning and obstacle avoidance, tested in Gazebo simulations.
XR · 3DVoice-driven CAD assemblyIIT Delhi · FSM Lab · 2025
onboarding time
−75%
mechanical fidelity
98%
voice → 3D CAD
Real time
Simulated
Tap a command
Pipeline
voice → Flask → Unity
Engine
Unity · C#
An industrial training toolkit in Unity and C#: spoken commands drive the 3D CAD models, and the physics matches the real hardware. Watch the toolkit

Stations on the traverse Experience

6 roles · 2 schools

  1. Next

    Next station: your team?

    Thirty minutes is enough to take a reading.

    Book a call

  2. Present

    Current

    AI Engineer Intern

    Hive Pro, Inc.

    Dubai, United Arab EmiratesRemote

    • Shipped production AI platforms: an ISMS portal, customer onboarding and sales-outreach agents
    • Human-in-the-loop by design: approval flows, RBAC, tenant isolation, policy guardrails
    • Agent pipelines on Claude, LangGraph and RAG, wired to HubSpot, Slack, Gmail, Fathom and NATS
    • Built and deployed production-grade AI platforms, including an ISMS portal, customer onboarding and sales outreach agentic platforms.
    • Designed human-in-the-loop AI systems with approval workflows, RBAC, tenant isolation, and policy-based guardrails to ensure safe and controlled automation.
    • Engineered LLM/agent pipelines integrating Claude, LangGraph, RAG, embeddings, and document processing, with external systems including HubSpot, Slack, Gmail, Fathom, and NATS.
    • Hardened production systems through security and risk audits covering authentication, authorization, API protection, secret management, data isolation, AI grounding, concurrency, and automated testing.
    • Claude
    • LangGraph
    • RAG
    • Embeddings
    • HubSpot
    • Slack
    • Gmail
    • Fathom
    • NATS

  3. 5 months

    AI Engineer Intern

    Rainwire Solutions Pvt. Ltd. · WARDEC – Indian Army

    New DelhiOn-site

    parameter LLM
    8B
    accuracy
    88–92%
    • Fine-tuned an 8B-parameter LLM on in-house datasets for 3-way classification, achieving 88–92% accuracy.
    • Built a secure mobile application for military personnel (soldiers and brigadiers) under NDA.
    • Developed a data acquisition pipeline using Playwright and Selenium, automating large-scale scraping and preprocessing.
    • Built an AI-based deepfake detection pipeline for videos, images, text and URLs to detect manipulated media.
    • Contributed to backend development and deployment of AI services on secure government infrastructure.
    • LLM fine-tuning
    • Playwright
    • Selenium
    • Deepfake detection

    Anything beyond this is classified.

  4. 3 months

    AI Engineer Intern

    Scrini AI Pvt. Ltd.

    Remote

    profiles searched
    10K+
    retrieval
    < 50 ms
    • Semantic resume search over 10K+ profiles with sub-50 ms retrieval (HNSW)
    • Hybrid retrieval with cross-encoder reranking for relevance
    • Shipped as 3-tier microservices with Docker Compose
    • Architected an AI-powered resume search engine using Python, FastAPI, and Qdrant, enabling semantic matching across 10K+ candidate profiles with sub-50ms retrieval latency via HNSW indexing.
    • Designed an asynchronous NLP pipeline with hybrid retrieval and cross-encoder reranking, combining vector similarity (cosine), metadata filtering, and deep semantic ranking for improved relevance.
    • Deployed a scalable 3-tier microservices architecture (FastAPI backend, Streamlit UI, vector database) using Docker and Docker Compose, ensuring production-ready orchestration and modularity.
    • Python
    • FastAPI
    • Qdrant
    • HNSW
    • Cross-encoders
    • Streamlit
    • Docker

  5. Present

    Current

    Technical Mentor (VR, Robotics & AI)

    Maker’s Asylum

    Remote

    students mentored
    250+
    cohorts
    3
    • Mentored 250+ students in robotics hardware and computer vision
    • Designed a scalable AI/ML curriculum for 3 cohorts
    • Arduino and AI workshops for underserved students (25+ participants)
    • Mentored 250+ students in robotics hardware and computer vision; designed a scalable AI/ML curriculum for 3 cohorts.
    • Empowered underserved students by conducting workshops on Arduino and AI technologies among 25+ participants.
    • Computer Vision
    • Arduino
    • Robotics
    • Teaching

  6. 4 months

    Extended Reality (XR) Intern

    Indian Institute of Technology, Delhi · FSM Lab

    New DelhiOn-site

    onboarding time
    −75%
    mechanical fidelity
    98%
    • Unity + C# industrial training toolkit: onboarding time cut by 75%
    • Voice commands synced to 3D CAD models in real time (Python Flask)
    • Physics-accurate scripts at 98% mechanical fidelity to the real hardware
    • Architected an interactive industrial training toolkit in Unity and C#, reducing employee onboarding time by 75%.
    • Developed a real-time AI integration pipeline using Python Flask to sync voice commands with 3D CAD models.
    • Engineered physics-accurate scripts, achieving 98% mechanical fidelity compared to physical hardware.
    • Unity
    • C#
    • Python
    • Flask
    • CAD
    Watch the toolkit

  7. Present

    Current

    Technical Team Member

    A.T.O.M Robotics Lab

    New DelhiOn-site

    • Autonomous GPS navigation for a Mars Rover prototype in ROS2 and Python
    • LiDAR + IMU sensor fusion for path planning and obstacle avoidance, tested in Gazebo
    • Built an autonomous GPS navigation system for a Mars Rover prototype using ROS2 and Python.
    • Implemented sensor fusion (LiDAR + IMU) for precise path planning and obstacle avoidance in Gazebo simulations.
    • ROS2
    • Python
    • LiDAR
    • IMU
    • Gazebo

  8. B.Tech in CSE (AI/ML)

    Maharaja Agrasen Institute of Technology

    New Delhi, IndiaEducation

    CGPA
    9.0+

  9. Class 12th · Class 10th

    St. Joseph’s Academy

    New Delhi, IndiaEducation

    Class 12th
    91.4%
    Class 10th
    94.4%

Projects

3 projects · live demos

01 / 03

Project 01Android · Full stack · AI/ML

SETU

GPS-denied egomotion tracking

Offline navigation that keeps tracking a vehicle through GPS outages, using the phone’s own sensors and maps.

207 KBon-device model

  • 1–5 ms inference
  • 227 Hz sensor pipeline
  • 84,000+ parameters
Built with
  • C++20
  • Kotlin
  • Android
  • PyTorch
  • TensorFlow Lite
  • Eigen
  • GNSS/IMU
  • RI-EKF
  • MapLibre
GNSS-DENIED UNDERPASSGNSSANDETAIL A · ON-DEVICE PIPELINEIMUaccel · gyroSVOspectralCNN-GRUvelocity modelRI-EKFfusionMAPMapLibre227 Hzsensor pipeline3207 KBon-device model184,000+parameters41–5 msinference2
Website / APK

How SETU works

Problem
GPS drops out in GNSS-denied environments, and navigation has to bridge the outage offline.
What he built
  • An offline navigation system for GNSS-denied environments, combining inertial sensing, road constraints, and map-based localization to bridge GPS outages.
  • A Spectral Velocity Odometry (SVO) pipeline that extracts wheel-rotation frequencies from smartphone accelerometer vibrations to estimate vehicle speed without double-integrating acceleration.
  • An on-device CNN-GRU velocity model trained on real vehicle recordings — 207 KB, 84,000+ parameters, 1–5 ms inference on mobile devices.
  • A high-rate sensor pipeline processing real driving data at up to 227 Hz, with offline maps, continuous GNSS-to-dead-reckoning handover, and deterministic replay for validation.
Project 02Multimodal ML

Product Pricing Prediction

Amazon ML Challenge

Predicts a product’s price from its listing text and its image.

Top 20%global leaderboard

  • 21.06% SMAPE
  • 150,000+ products
  • 15% better than baseline
Built with
  • Python
  • PyTorch
  • scikit-learn
  • Pandas
  • NumPy
  • Transformers (BERT)
  • LightGBM
  • OpenCV
BERTtext embeddingRESNETimage embeddingfusedXGBOOSTLIGHTGBMENSEMBLEPRICE150,000+products3Top 20%global leaderboard121.06%SMAPE2baseline modelsensemble · lower error15%better than baseline4

How Product Pricing Prediction works

Problem
Predicting a product’s price from its text and its image.
What he built
  • A multimodal pricing pipeline processing 150,000+ products using text (BERT) and image (ResNet) embeddings.
  • 21.06% SMAPE by ensembling XGBoost and LightGBM models, outperforming baseline models by 15%.
  • Top 20% placement on the global leaderboard.
Project 03Web3 · Full stack

TokenForge

Aptos blockchain launchpad

A token launchpad on Aptos that writes the Move contract for you and locks the liquidity.

90%supply locked in liquidity pools

Built with
  • Next.js
  • React
  • Node.js
  • Supabase
  • Aptos
  • Move
MOVE MODULEauto-generatedLAKSHYACOINLIQUIDITY POOLS · LOCKEDSUPPLY90%supply locked in liquidity pools1
Watch the demo

How TokenForge works

What he built
  • A full-stack token launchpad with automated Move contract generation, locking 90% of supply in liquidity pools.
  • Launched a coin named “Lakshyacoin” — real users bought it with the Petra wallet, with all coin details on the platform.

Rendezvous Book a call

30 min · Cal.com

AI agents, full-stack products and 3D websites, built and shipped to production.

  1. 1Pick a day
  2. 2Pick a time
  3. 3The invite lands in your inbox
Pick a time30 minOpen on cal.com

Loading his calendar…

Prefer to write?

10jainlakshya@gmail.com

The legend Tech stack

30 tools · 5 layers

Hover or tap a tool to see where he used it.places on this page that use it · 24 of 36 traced

AI / LLM9

Frameworks & tools9

Languages6

Core concepts6

Ways of working6

Survey log GitHub activity

Contribution terrain, waiting for GitHub.FebMarAprMayJunJulAugSepOctNovDecJan 26MonWedFri
Contributions, last 12 months
—
Active days
—
Busiest day
—
Period
—
Source
—
Surveyed
—

Latest commitsreading…

  6. Reading the field book…

Benchmarks Achievements

6 highlights