I learned from a financial-news deduplication failure that a strong average metric can coexist with a severe production failure. I now inspect worst cases alongside summary metrics, and I treat cluster-level behavior as something pairwise precision alone cannot establish.
In my financial-news deduplication system, individually correct pairwise matches chained 208 unrelated articles into one cluster, a failure pairwise precision could not reveal.
I changed the financial-news clustering method from single linkage to complete linkage after finding the oversized false cluster.
I now print the worst case beside every summary number because an average can make things look fine while hiding this kind of bug.
I build systems with a particular focus on making their outputs and behavior checkable. My projects apply that principle to meeting notes, browser-based bug reproduction, market calls, architecture evaluation, and seat reservations.
I spend longer building the thing that checks the thing I build.
Verbatim is a Hinglish meeting notetaker with a cite-or-drop rule: every action item must point to a timestamp or it is deleted.
Witness drives a real browser to reproduce a human bug report and will not say CONFIRMED until checks trip cleanly three times.
Hindsight seals each market call before the outcome exists.
ArchForge has a deterministic verifier that scores the architecture, so the model is not marking its own homework.
Boxoffice makes the seat check and the seat lock the same SQL statement.
My contrarian opinion is that a benchmark number tells you about the benchmark, not that the system works.
In Hindsight, my clustering precision read 1.00 while a 208-article false cluster sat in the output.
Complete linkage took the largest cluster to 2.
After recalibrating a threshold I had tuned on label-file IDF, false matches on random pairs fell from 13 percent to 0.16 percent over 20,000 pairs.
In Gauntlet, prompt-injection detectors report 99 percent on their own distribution; I held out whole attack families and scored against 662 real human-written injections, where AUC came back 0.56.
Canonicalisation plus sentence windowing took Gauntlet's AUC to 0.81; swapping in a fine-tuned transformer did nothing.
People come to me to make model output checkable and to drop into a large unfamiliar codebase to ship a small, correct fix.
I made 16 pull requests to Apache DataFusion, with 11 merged, mostly for Substrait plan serialization.
I made 20 pull requests to Apache Airflow, with 4 merged.
- Other open-source projects I have contributed to include AsyncAPI, RUXAILAB, and Kolibri.
I am weak in C and Java.
My CGPA is 7.5 out of 10.
My evaluation work uses my own benchmarks rather than published research.
I am pursuing a BTech in Computer Science and AI at Newton School of Technology, Rishihood University, from 2024 to 2028.
I was a Software Engineer Intern at Klariti Learning Innovations from May 2025 to July 2025.
The captured description of my GenAI trainer role at Octa Learning ends mid-entry, so the remainder of that role description is unavailable.
My headline is: Full-Stack & AI Engineer | Building LLM-powered products (RAG, Agents, System Design) | React · Node.js · Python · PyTorch | Built ArchForge (98% recall AI system-design engine) | CS+AI @ NST ’28.
I'm a Computer Science & AI undergrad (Newton School of Technology, '28) focused on Applied AI and Full-Stack Development.
I build production-ready AI systems that combine LLMs with scalable backend engineering, taking ideas from first commit to deployment. My work spans RAG systems, AI agents, system design, backend APIs, and modern web apps, with hands-on ML in PyTorch. I work at the intersection of AI, backend engineering, and product, turning ambiguous ideas into shipped, reliable software. I am open to internships and early-career roles in Applied AI, AI Engineering, and Full-Stack Engineering.
I built ArchForge, an AI system-design engine that turns plain-English prompts into production-grade architecture diagrams. It pairs LLM generation with a deterministic verifier that repairs illegal edges and missing auth/caching layers — ~98% recall against real architectures (Uber, Netflix, Instagram) and 100% across 10 held-out domains, covered by 75 tests.
I delivered 10+ production full-stack apps for clients, most live with real active users — owning delivery end-to-end from architecture to deployment.
I shipped features on a live ed-tech platform at Klariti Learning, cutting page load times ~30% and serving thousands of API requests/day.
I trained 20+ engineers on applied GenAI — LLMs, RAG, and prompt engineering — through hands-on, build-real-things projects.
My position is Full-Stack & AI Developer at Freelance (Jan 2025 to Present).
My position is GenAI Trainer at Octa Learning (Aug 2025 to Jul 2026), New Delhi.
My position is Software Engineer Intern at Klariti Learning (May 2025 to Jul 2025).
I studied for a Bachelor of Technology - BTech, Computer Science and Artificial Intelligence at Newton School of Technology (2024 to 2028).
I completed Intermediate (Class XII) at Delhi International Public School (2023 to 2024).
My technical skills include PyTorch, REST APIs, MongoDB, Artificial Intelligence (AI), GitHub, Git, Chart.js, Front-End Development, Web Applications, Web Development, Full-Stack Development, CryptoAPI, Software Packaging, Programming Languages, Tailwind CSS, Node.js, React.js, JavaScript, Cascading Style Sheets (CSS), HTML5, HTML, XHTML, Web Design, JavaScript Libraries, Website Building, Microsoft Power BI, Microsoft Excel, Microsoft PowerPoint, Scratch, Figma (Software), BrowserStack, and Stack. My other skills and strengths include Critical Thinking, Problem Solving, Decision-Making, Virtual Assistance, Educational Technology Integration, Computer Literacy, Market Mapping, Real-time Data Acquisition, Portfolio Management, LinkedIn, Team Collaboration, and Animation Design.
In my freelance work, I deliver production full-stack web applications for clients from requirements through deployment, design scalable backends and REST APIs, and serve as each client's technical point of contact. At Klariti Learning, I shipped features across a React and Node.js ed-tech platform, built APIs for course and user workflows, and worked with a cross-functional team. At Octa Learning, I trained engineers in applied Generative AI and designed hands-on projects integrating AI-powered applications. My contact email is namanjainpy@gmail.com.