01
Vision Heavy GUI Auth Agent
LLM-driven browser automation using Playwright + OmniParser, with self-correcting execution, safety guardrails, and human-in-the-loop fallbacks.
AI Systems Engineer · Agentic AI · LLM Systems · Computer Vis.
Nikhil Lamba
Building AI systems that reason, act, and operate in production.
I work on agentic AI, LLM systems, multi-agent platforms, and deep-learning systems — from research and experimentation to production. I also build useful web apps over weekends, ship them publicly on X, and share what I learn as an active open-source contributor in X.com's indie-hacking communities.
Recent: Founding AI Engineer @Adopt AI
60+agents in production
1,000+successful autonomous actions / month
9+enterprise clients
95%+reduction in production regressions
Where I've built and shipped.
-
Built a collaborative platform for marketing teams to engage and decide which version of storyboard is better to go for AI video production.
- Fine-tuned and benchmarked Qwen 3 LLM via SFT and RL (DPO, PPO).
- Architected an end-to-end agentic harness as per defined RBACs.
- Architected and scaled a collaborative environment based on unique usernames.
- Added proper RBAC functionalities for team members.
- Interactive versioning of text-to-video ideas with team-wide visibility.
-
Founding AI Engineer Adopt AI Mar 2025 — May 2025 Bengaluru, Karnataka, India · Agentic Platform & Enterprise Workflow Automation
Building a production multi-agent platform from the ground up for 9+ global enterprise clients — orchestration, tool calling, evaluation, and execution infrastructure. Led a team of 8 engineers.
- Multi-agent orchestration: architected a custom Workflow Descriptive Language (WDL) and an "Uber Agents" routing harness scaling to 60+ autonomous agents and 1,000+ successful actions/month.
- LLM-as-a-Judge evals: benchmarked routing accuracy, tool-calling precision, and safety boundaries — cutting regressions by more than 95%.
- Remote code execution: shipped RCE as a core architecture shift, unlocking agent-to-agent autonomous behavior.
- Enterprise integrations: zero-human-in-the-loop agents across QuickBooks, Gmail, Google Drive, Plaid, and US Bank APIs (98% automation) with human-in-the-loop guards for sensitive actions; recovered $1.8M+ in annual billing leakage.
- Self-healing SOQL: built an NL-to-SOQL compiler with a 3-loop fallback that corrects syntax and logic errors at runtime.
- DevEx: a CLI toolchain that cut deployment latency 85% and accelerated delivery 4x.
-
Tech Lead HCLTech Oct 2023 — Feb 2025 Chennai, India · Generative AI & Multimodal Computer Vision
- Generative document restoration: a CycleGAN pipeline on 15,000+ distorted words hitting 82% accuracy — the proof-of-concept secured a multi-year US enterprise contract.
- Interactive object extraction: cascaded Grounding DINO + Segment Anything Model (SAM) into a zero-shot system, cutting AWS EC2 memory use by 45% and eliminating memory leaks.
- Spatial 3D reconstruction: deployed COLMAP and InstantNGP for indoor reconstruction from monocular feeds with high-precision distance estimation.
-
AI & ML Intern Proxgy Apr 2022 — Jun 2022 Gurugram, India
- Built an automated vehicle-damage estimation framework using instance segmentation, optimizing edge inference for real-time insurance repair-cost estimation.
Systems I've designed and shipped.
Agentic systems, deep-learning pipelines, and the weekend web apps I build and publish in public.
02
Collaborative Image Editing for Storyboards
A collaborative image-editing workflow for storyboards — multiple creators shaping and iterating on frames together in real time.
03
Soccer Scene Understanding & 3D Volumetrics
3D scene estimation from monocular soccer broadcasts — novel videogame-derived datasets via GPU cycle alteration, a faster pipeline for monocular depth generation, and Detectron + OpenPose skeletal tracking.
04
Interactive Scene Segmentation, A Prompt Centric Approach
One-shot interactive segmentation: an end-to-end pipeline spanning prompt input, object detection, and pixel-level segmentation for image editing.
05
Object Tracking by Detection
YOLOv5 + DeepSORT and OpenCV trackers benchmarked head-to-head with quantitative accuracy metrics for real-time scenarios.
06
Synthetic Haze Generation and Tracker's Analysis
Synthetic haze dataset creation and validation of state-of-the-art trackers for autonomous vehicles and UAVs operating in degraded visibility.
07
Broken Image Reconstruction
Reconstruction of old, crushed, or broken documents using modified GAN weights — built on a custom dataset (none existed publicly) and a customized conditional GAN architecture.
08
Virus Lungs Classification App
A Flask web app with a CNN that classifies chest X-rays as Viral Pneumonia, Bacterial Pneumonia, or Normal — with the infected region highlighted.
09
Contactless Disease Prediction
A quarantine-era backend that maps two self-reported symptoms to five likely diseases, ranked by severity, using an ANN.
Papers & patents.
Journal · Springer · 2024
Single Object Tracking under Variable Haze
Multimedia Tools and Applications
Benchmarked YOLOv7, SORT, and SiamBAN under synthetic haze degradation.
Conference · ICDAIC · 2023
Faster Pipeline for Depth Generation of Soccer Players
N. Lamba, S. Singh, A. Khosla
Faster pipeline for monocular depth generation of soccer players, presented at the 2023 ICDAIC conference.
Patent · Indian Patent Office · 2021
SARS-CoV-2 Classification via Chest Radiograph
Application No. 202111030911A
Deep-learning classification of chest radiographs.
Shipped in public, with the lessons attached.
An engineer at the boundary of research and production.
I'm an AI engineer focused on building intelligent systems that operate reliably in the real world. My work spans agentic AI, LLM systems, computer vision, and deep learning — across production multi-agent platforms, enterprise automation, multimodal systems, and published research.
I enjoy working at the boundary between research and engineering: taking ideas that work in a notebook and turning them into systems people can actually use. Outside work, I build small AI products and engineering experiments, ship them publicly on X, and contribute to X.com's indie-hacking and open-source communities.
Technical interests
Core AI
Tools & Platforms
Formal training.
-
M.Tech — Artificial Intelligence NIT Jalandhar Aug 2021 — Aug 2023 Jalandhar, India
CGPA 8.86/10 · Specialization: Deep Learning, Computer Vision & Advanced AI Algorithms.
-
B.Tech — Computer Science & Engineering SRM University, Delhi-NCR Aug 2017 — Jul 2021 Sonepat, India
CGPA 8.67/10 · Gold Medalist (Ranked 1st in Department overall).