-Newton_School_of_Technology-1F1147?style=for-the-badge&logo=academia&logoColor=white){kind=icon}
I’m Lakshya Gupta, a Software Engineer and AI/ML builder focused on shipping production-grade systems with strong software engineering fundamentals. I’ve been coding for 10+ years, and I care deeply about clean architecture, debugging discipline, performance, evaluation rigor, and building products that work at real-world scale.
My work sits at the intersection of software engineering, search, retrieval systems, LLM infrastructure, and full-stack product development. I’ve built and deployed systems used by 12,000+ users across India, Vietnam, Indonesia, and the Philippines, spanning agentic RAG, LLM evaluation and red-teaming, learning-to-rank search, AI-powered developer tooling, and education products. I approach AI as an engineering problem: measurable, observable, testable, and production-bound.
From a product engineering standpoint, I enjoy taking systems end-to-end - from architecture and backend services to frontend UX, deployment pipelines, and post-launch evaluation. My stack typically blends Python, TypeScript, FastAPI, React, PostgreSQL, vector databases, and modern LLM tooling, with a strong emphasis on retrieval quality, guardrails, experimentation, and reliability.
- Software Engineering Internships - backend, full-stack, platform, infra-adjacent roles
- AI/ML Engineering Internships - LLM systems, evaluation, search, RAG, MLOps
- Open Source Collaborations - developer tools, AI tooling, education tech, infra
- Research / Applied Engineering Projects - responsible AI, retrieval, ranking, model evaluation
| Domain | Proficiency | Details |
|---|---|---|
| Retrieval-Augmented Generation (RAG) | Advanced | Architected agentic and hybrid RAG systems with BM25 + dense retrieval + reranking + self-reflection loops, grounded citations, and quality-gated response generation. |
| LLM Evaluation & Red-Teaming | Advanced | Built evaluation pipelines with DeepEval, LangSmith, MART-style mutation testing, exportable audit reports, and guardrail validation across multiple failure classes. |
| Search & Ranking Systems | Advanced | Built LambdaMART learning-to-rank models on 2M+ query-document pairs, applied IPW propensity correction, and evaluated via NDCG / MAP / MRR with shadow deploy + A/B testing. |
| LLM Application Engineering | Advanced | Production experience with FastAPI + React AI products, multi-model orchestration, prompt design, retrieval evaluation, tracing, and performance optimization. |
| Responsible AI / Hallucination Mitigation | Advanced | Strong focus on grounded generation, verifiable citations, policy guardrails, red-teaming, and explainability for production AI systems. |
| Data / ML Tooling | Advanced | Comfortable across Pandas, NumPy, SciPy, scikit-learn, XGBoost, Gradio, embeddings pipelines, experiment analysis, and evaluation instrumentation. |
| Knowledge Graphs & Developer AI | Intermediate → Advanced | Built code understanding systems using AST-level embeddings + Neo4j traversal + commit metadata fusion for explainable AI-assisted developer workflows. |
| Full-Stack Product Engineering | Advanced | Experience shipping products across backend services, frontend UX, deployment, observability, auth, storage, and real-world iteration loops. |
Agentic Hybrid RAG Platform
Production-grade agentic RAG platform designed for verifiable answers, retrieval quality, and measurable reliability. The system combines BM25 sparse retrieval, ChromaDB dense retrieval, Cohere reranking, and an autonomous self-reflection loop that reformulates queries until response quality criteria are satisfied. Built as a full-stack product with a FastAPI backend and React frontend, and instrumented with Ragas-based evaluation for faithfulness and answer relevancy.
| Category | Details |
|---|---|
| Stack | Python, FastAPI, TypeScript, React, LangChain, LangGraph, ChromaDB, Cohere, Ragas, OpenRouter |
| Scale | Production-ready full-stack deployment with multi-model execution |
| Performance | 92% context retrieval accuracy, sub-second query processing, exceeded 0.80 faithfulness and 0.75 answer relevancy thresholds |
| Security | Grounded responses with granular source citations and evaluation-driven quality gates |
| Impact | Built as a verifiable AI system focused on reducing hallucinations and improving trust in retrieval-heavy workflows |
| Repository | Add repository link here |
Engineering Notes
- Designed for fully grounded responses with source-level transparency
- Introduced a self-correcting retrieval loop to improve answer quality before final generation
- Focused on production deployment reliability, measurable evaluation, and modular architecture
- Represents my approach to AI engineering: retrieval quality + evaluation rigor + shipping discipline
LLM Evaluation & Red-Teaming Platform
A deployed LLM reliability platform for adversarial testing, policy validation, and failure analysis. The platform evaluates models across hallucination, safety, policy, and prompt attack scenarios, combining DeepEval metrics, LangSmith tracing, and mutation-based red-teaming into a single workflow with exportable audit reports.
| Category | Details |
|---|---|
| Stack | Python, FastAPI, TypeScript, SQLite, DeepEval, LangSmith |
| Scale | Tested 528 prompts across 6 failure categories |
| Performance | Identified 20.5% baseline violation rate and 4.5% hallucination rate; reduced violations to 0.0% after guardrail mitigation |
| Security | Policy-oriented evaluation workflows, red-team prompts, and audit-ready exports |
| Impact | Enabled structured, measurable evaluation of LLM behavior rather than relying on anecdotal testing |
| Repository | Add repository link here |
Engineering Notes
- Built to make LLM reliability an engineering workflow, not a manual QA task
- Supports JSON / CSV / Markdown / PDF audit report generation
- Useful for prompt hardening, regression checks, and policy validation
- Reinforced my interest in responsible AI, model evaluation, and guardrailed deployment
Search Ranking Model
A large-scale learning-to-rank system built to improve retrieval quality over lexical baselines. The model uses LambdaMART over 2M+ query-document pairs and applies inverse propensity weighting (IPW) to correct for position bias in click data. The project was evaluated using NDCG@1/5/10, MAP, and MRR, and validated through shadow deployment + A/B testing.
| Category | Details |
|---|---|
| Stack | Python, Pandas, NumPy, XGBoost, scikit-learn, BM25, SciPy, Pytest, Gradio |
| Scale | 2M+ query-document pairs |
| Performance | Improved NDCG@10 by 57% over BM25 baseline |
| Security | Evaluation-first design with testing and shadow validation before rollout |
| Impact | Demonstrated how ranking quality can be improved using robust offline metrics and bias-corrected learning |
| Repository | Add repository link here |
Engineering Notes
- Strongly focused on search relevance engineering, not just model training
- Bias correction using IPW made the click data more reliable for ranking optimization
- Validated through ablation analysis and deployment-aware testing
- This project reflects my interest in search infrastructure, ranking systems, and experimentation
Virtual Robos - Browser-Based Robotics Playground
A zero-install robotics education platform built for accessibility and real-time interactivity. The system uses OpenCV.js, WebRTC, and browser-native rendering to simulate robotics interactions, gesture-driven controls, and modular experimentation without requiring physical hardware.
| Category | Details |
|---|---|
| Stack | JavaScript, OpenCV.js, WebRTC, HTML5 Canvas |
| Scale | 7,800+ users |
| Performance | <50ms latency at 30fps with real-time interaction optimizations |
| Security | Client-side browser-first architecture with modular open-source workflows |
| Impact | Makes robotics experimentation accessible to under-resourced schools and learners without hardware |
| Repository | Add repository link here |
Engineering Notes
- Uses HSV filtering, contour detection, gesture logic, and real-time browser optimization
- Built around education access and zero-install usability
- Also served as a meaningful open-source collaboration and contributor workflow experience
Mar 2026 - Present · Remote
Driving AI/ML initiatives within the developer community while contributing to engineering workshops, collaborative learning environments, and technical problem-solving programs. This role sits at the intersection of community leadership, software engineering practice, and AI education, with a focus on helping students build stronger engineering intuition around modern AI systems and tooling.
Scope of Work
- Led AI/ML-focused community efforts and supported technical workshops around modern AI tooling and engineering workflows
- Applied software engineering best practices across collaborative build sessions, debugging workflows, and open-source-style environments
- Ranked Top 15,000 / 150,000 globally at Google Big Code 2026, reflecting strong algorithmic problem-solving and engineering discipline
- Contributed to a culture of practical learning around LLM systems, AI tooling, and developer productivity
Skills
Python AI/ML Developer Community Problem Solving Debugging Open Source Technical Leadership
Oct 2025 - Apr 2026 · Remote
Contributed across 10+ repositories in AI/ML and developer tooling, collaborating through professional Git workflows and maintainer review processes. This work sharpened my ability to understand unfamiliar codebases, ship meaningful contributions, and operate within collaborative engineering environments.
Scope of Work
- Contributed to 10+ repositories spanning AI/ML, tooling, and developer workflows
- Worked through full Git-based contribution loops: forking, branching, rebasing, code review, and CI-backed merge flows
- Ranked Top 10,000 / 98,000+ at Hacktoberfest 2025
- Collaborated across global open-source programs including OSCG 2026 and GSSoC 2026
- Strengthened engineering fundamentals around maintainability, review discipline, and contributor communication
Skills
Git GitHub Code Review CI/CD Open Source Python Developer Tooling
Oct 2025 - Present · Pune, India
Mentoring peers in AI/ML engineering through project guidance, concept reinforcement, and practical roadmap support. My focus is on helping students move beyond tutorial-level knowledge toward real engineering execution, especially in LLM systems, project architecture, and applied ML workflows.
Scope of Work
- Guided peers through project design, experimentation, and implementation in AI/ML
- Helped mentees build stronger understanding of RAG, model evaluation, data pipelines, and engineering trade-offs
- Supported hands-on learning through project feedback and structured problem-solving
- Multiple mentees secured AI/ML opportunities through practical portfolio-building support
Skills
Mentorship AI/ML RAG Project Guidance Engineering Communication
| Recognition | Details |
|---|---|
| HPAIR 2026 - Selected Delegate | Chosen for Harvard’s flagship Asia-Pacific conference in Hanoi, Vietnam, representing a high-potential global student cohort across 50+ countries. |
| Google Big Code 2026 | Ranked Top 15,000 / 150,000 globally in a competitive debugging and algorithmic engineering challenge. |
| Hacktoberfest 2025 | Ranked Top 10,000 / 98,000+ through open-source contributions across engineering repositories. |
| Hackathon Judge - Agentic AI Track | Evaluated 100+ teams across two events at Newton School of Technology, with recognition for LLM and applied AI domain expertise. |
| Founder - Sparker Kit | Won a 10-day entrepreneurship hackathon at age 16; generated ₹40K revenue and ₹20K profit. |
| Director General - IIT Delhi MUN 2024 | Led multi-committee operations and produced conflict-resolution reports adopted across 3 committees. |
