Saksham
Jain.
Designing AI systems that learn, reason, and act.
AI Engineer · Research Assistant · Robotics Builder. Working at the intersection of foundation models, vision-language systems, robotics and data engineering — turning open research questions into shipped systems.
An engineer–researcher building toward general-purpose intelligence.
I'm Saksham Jain — a Bachelor of Advanced Computing student at the University of Sydney (majoring in Computational Data Science and Computer Science).
My work spans large language models, vision-language robotics, meta-learning, and full-stack data systems. I've worked as a Research Assistant at NUS on automatic LLM strategy selection, as an AI Engineer Intern at CruvixAI building language-grounded robotic manipulation, and as a Technology Intern at Deloitte architecting real-time manufacturing dashboards across four factories.
Beyond formal roles, I publish peer-reviewed research, ship side products (most recently ThinkPost, an AI LinkedIn-content SaaS), and compete — Top 4 at the SparkSoc × SoftSoc Hackathon and Top 4 at the SUBAA × SUDATA Datathon. I'm most alive when an open research question turns into a deployed system.
Building AI that reasons about new problems and acts in the real world — not just predicts on benchmarks.
Research-first, engineering-honest. Tight feedback loops between papers, code and the systems they run on.
Where I've built and shipped.
Industry, research and leadership roles across AI, robotics and data systems.
AI Engineer Intern
- Building a PyBullet simulation environment featuring a robotic arm performing pick-and-place tasks on cluttered shelves, integrating open-source vision-language models (VLMs) from HuggingFace to ground natural-language instructions to specific objects.
- Designing a systematic evaluation framework across instruction complexity levels and clutter densities to identify VLM failure modes — including similar-object confusion, vague-instruction sensitivity, and partial-occlusion robustness.
- Authoring a research paper documenting grounding-accuracy benchmarks and failure analysis, contributing to applied robotics and language grounding at the intersection of NLP and computer vision.
Research Assistant
- Investigating whether a meta-learning system can automatically select the optimal LLM prompting strategy (zero-shot, few-shot, RAG) for a given NLP task by learning meta-features that characterise task complexity, domain, and data availability.
- Benchmarked three prompting strategies across 7 sentiment-analysis datasets (IMDB, SST-2, Yelp, Amazon, Emotion, Tweet, Rotten Tomatoes) using macro F1; extracted meta-features including text length, vocabulary size, training-set size, class entropy, and label-imbalance ratio.
- Trained a Random Forest meta-learner on the extracted meta-features, achieving meaningful accuracy improvements in strategy selection over baseline trial-and-error approaches.
Technology Intern
- Architected and deployed a real-time manufacturing dashboard monitoring machine performance across 4 factories in Python, enabling live fault detection and reducing manual reporting overhead.
- Built end-to-end data-transformation pipelines that ingested multi-source JSON telemetry, standardised formats, and surfaced clean, actionable metrics for operations teams.
- Authored a formal project proposal covering dashboard functionality, UI/UX design, role-based access control, development estimates, and a phased delivery timeline — approved by senior stakeholders.
Executive Member
- Represented USYD in RoboWars and the Autonomous Robotics Challenge, building integrated robotic systems using ROS, Arduino, and Raspberry Pi with custom sensor-fusion and control algorithms.
- Engineered the USRC × ANT61 Hackathon (100+ participants, 48 hours) — designed a real-time satellite-threat-detection challenge using React.js, TypeScript, and WebSockets, and judged all final submissions.
Executive Member
- Contributed to Ignite, a 200+ participant flagship event covering AI, MVP building, pitching, and business-model workshops — managed logistics, speaker coordination, and on-day execution.
- Co-organised the StartupLink × Lyra × Build Club Hackathon, connecting students and mentors to develop solutions aligned with UN SDGs 3, 10, and 11.
Two research threads, one north star.
Making intelligent systems that pick the right strategy and ground language into real-world action.
Meta-Learning for Automatic LLM Strategy Selection
Benchmarking zero-shot, few-shot and RAG across 7 sentiment datasets (IMDB, SST-2, Yelp, Amazon, Emotion, Tweet, Rotten Tomatoes), then training a Random Forest meta-learner over task meta-features — text length, vocab size, training-set size, class entropy, label-imbalance ratio — to predict the optimal strategy.
Result: measurable accuracy lift over trial-and-error baselines, reducing the manual overhead of deploying LLMs to new tasks.
Language-Conditioned Robotic Manipulation with VLMs
A PyBullet arm performs pick-and-place on cluttered shelves while open-source vision-language models ground natural-language instructions — "pick the red bottle on the left" — to specific objects in the scene.
Systematic evaluation across instruction complexity and clutter density exposes failure modes: similar-object confusion, vague-instruction sensitivity, partial-occlusion robustness.
Systems I've designed, built and shipped.
A working portfolio across AI products, robotics, full-stack and data engineering.
ThinkPost
AI-Powered LinkedIn Content SaaS · Jan 2026 — Present
- Building a full-stack LinkedIn content-creation platform in Next.js, Supabase, and Groq API — enabling creators to capture ideas from text, YouTube transcripts, PDFs, and voice memos, and generate structured posts.
- Implemented AI post generation with multiple content formats (Personal Story, Data Insight, Hot Take, How-To), hook-strength scoring, a content calendar, draft management, and post analytics.
- Architected Supabase-backed auth, a template system, and an admin dashboard; deployed on Vercel with a complete marketing site including features, pricing, blog, and about pages.
Autonomous Tracked Robot
Raspberry Pi-Powered Rover · Sep 2025 — Nov 2025
- Led end-to-end development of a Raspberry Pi-powered tracked robot: chassis assembly, sensor integration (ultrasonic, infrared, camera), and Python-based control systems.
- Implemented autonomous navigation, real-time obstacle avoidance, and remote operation via custom algorithms and sensor fusion; mentored teammates in computer vision and autonomy.
Strata Management Hub
Full-Stack Property Platform · Mar 2025 — Jun 2025
- Built a full-stack property-management platform (PHP, Neon PostgreSQL, Vercel, AWS Lambda / SNS / API Gateway) supporting resident directories, levy tracking, maintenance workflows, and automated notice generation.
- Delivered multi-page database integrations, session-based authentication, and serverless backend workflows that cut manual administrative steps for building managers.
Manufacturing Ops Dashboard
Real-Time Multi-Factory Monitoring · Jul 2025 — Dec 2025
- Real-time dashboard monitoring machine performance across 4 factories — live fault detection, ingestion of multi-source JSON telemetry, and clean operational metrics.
- Authored a stakeholder-approved project proposal covering UI/UX, RBAC, development estimates, and a phased delivery timeline.
Mosaic
Creative Professional Networking · Sep 2025 · Hackathon
- Built Mosaic — a creative professional networking platform blending portfolios with social-discovery features.
- React + TailwindCSS frontend and Node.js backend, designed and shipped within a 24-hour hackathon window.
VLM Manipulation Pipeline
Language → Action for Robot Arms · Jun 2026 — Present
- PyBullet simulation of a robot arm performing language-conditioned pick-and-place on cluttered shelves.
- Open-source HuggingFace VLMs ground natural-language instructions to specific objects; systematic eval across complexity and clutter density.
Recognition for work that shipped.
Top 4 / 200+ · SUBAA × SUDATA Datathon 2025
Ranked Top 4 in a high-difficulty logistics datathon with sparse, synthetic, and non-continuous data. Engineered domain-driven features (Fuel Efficiency Index, Congestion Stress, Logistics Resilience) and applied TabTransformers and XGBoost pipelines.
Top 4 / 150+ · SparkSoc × SoftSoc Hackathon
Built Mosaic — a creative professional networking platform blending portfolios with social-discovery features. Developed in React + TailwindCSS (frontend) and Node.js (backend) within 24 hours.
UG High Honour Roll 2025 · USYD
Awarded by the School of Computer Science for achieving a High Distinction average in Computer Science — recognition for academic excellence in 2025.
Dalyell Scholar · University of Sydney
Selected for USYD's elite Dalyell Scholars program, recognising students with outstanding academic achievement and offering accelerated, enriched coursework.
NVIDIA DLI · Getting Started with Deep Learning
Trained CNNs from scratch in PyTorch (MNIST, ASL); applied data augmentation, transfer learning, and fine-tuning. Built NLP pipelines using BERT for tokenization and question-answering.
NVIDIA DLI · Rapid Application Dev. with LLMs
Built LLM-powered apps with HuggingFace Transformers (semantic embeddings, zero-shot classification, text generation) and multimodal pipelines (CLIP, diffusion models). Orchestrated agents via LangChain.
The toolkit I build with.
Languages
Data & ML
Tools & Platforms
Mathematics
Coursework, scholarships, cohort recognition.
University of Sydney
Applied Data Science · Software Engineering & Design · Network Platforms · Operating Systems · Discrete Mathematics · Programming Languages & Techniques · Computer Architecture
National University of Singapore
Algorithms · Machine Learning · Databases · Statistical Learning · Data Engineering
What I'm doing right now.
Updated June 2026 · Jaipur → Singapore → Sydney
- Building VLM-powered robotic manipulation pipelines @ CruvixAI — language-conditioned pick-and-place on cluttered shelves.
- Running meta-learning experiments @ NUS — Random Forest meta-learner over 7 NLP datasets to predict optimal LLM strategy.
- Shipping ThinkPost — an AI LinkedIn-content SaaS on Next.js + Supabase + Groq, with calendar, analytics, and admin dashboard.
- Studying advanced ML theory at NUS — optimisation, generalisation, statistical learning.
- Open to research collaborations, internships in AI/robotics, and shipping side products.
Let's build something that reasons and acts.
I'm open to research collaborations, AI/robotics internships, and conversations about frontier ML, agentic systems, and what intelligent machines should do next. Email is fastest — I usually reply within a day.
jain.saksham2006@gmail.com →