Home › Tracks › Applied ML & Model Engineering
12-week elite track
Take a base model and make it yours.
Go from neural-net first principles to shipping adapted models: transformer pretraining intuition, supervised fine-tuning, parameter-efficient methods (LoRA/QLoRA), preference optimization (RLHF/DPO), distillation, and rigorous evaluation. You leave able to own a model-customization pipeline end to end. Bridges to Machine Learning, Linear Algebra, and Statistics.
Week by week
Every week unlocks the next. Concepts route you to free, world-class material; projects turn that knowledge into something deployed.
Week 1
Build a neural net and autograd by hand. Gradients, the chain rule, and what 'training' actually computes — no framework magic until you have earned the abstraction.
Bridges to Calculus & Linear Algebra — gradients, the chain rule, and vector spacesBuilds on: nothing — start here
Week 2
Optimization that actually converges: SGD and Adam, regularization, normalization, initialization, and the failure modes (vanishing gradients, overfitting) every practitioner must recognize.
Bridges to Machine Learning — optimization, generalization, and the bias-variance tradeoffBuilds on: Neural Networks & Backpropagation from Scratch
Week 3
Why attention replaced recurrence, and what pretraining a language model on a corpus actually optimizes. Tokenization, the training objective, and scaling laws.
Bridges to Machine Learning — sequence modeling and representation learningBuilds on: Deep Learning Foundations
Week 4
Model quality is data quality. Building, filtering, and deduplicating datasets; instruction-data curation; contamination; and why most fine-tuning failures are data failures.
Bridges to Databases — data cleaning, deduplication, and ETL pipelinesBuilds on: Transformers, Attention & Pretraining
Week 6
Adapt a base model to a task or style with SFT: instruction tuning, hyperparameters that matter, overfitting and catastrophic forgetting, and measuring whether it worked.
Bridges to Machine Learning — transfer learning and supervised trainingBuilds on: Training Data Engineering
Week 7
Adapt billion-parameter models on a single GPU. Low-rank adaptation, QLoRA's 4-bit base plus adapters, and the cost-versus-quality math that makes customization affordable.
Bridges to Linear Algebra — matrix rank, decomposition, and low-rank approximationBuilds on: Supervised Fine-Tuning
Week 8
An enterprise mandate: take a base open-weight model and adapt it into a specialist for a real domain you choose (legal, medical, code, support), then ship it as a launched product. Own the whole pipeline: build and clean the training dataset, run parameter-efficient fine-tuning (LoRA or QLoRA), and prove on a held-out, contamination-controlled evaluation that the adapted model beats the base model on the target task. The deliverable is not a notebook — it is a directly deployable, hyperscalable product: the fine-tuned model served behind a real public API with a hyper-usable demo UI, autoscaling, CI/CD, observability, security, and full marketing (landing page, pitch, demo) so a domain user can try it and a buyer can evaluate it. We do not accept 'it seems better' — bring the numbers, and ship it as a real product.
Why it matters: Domain-adapted models are how companies turn a generic LLM into a defensible product, and most fine-tuning projects fail on data discipline. A builder who can run a clean, evaluated, reproducible fine-tuning pipeline is directly deployable as an ML Engineer or Model Engineer, a frontier role at the ₹1-crore tier.
A publicly hosted product with a stable URL and a hyper-usable demo UI, plus a public repo and a published model card: the data pipeline, the QLoRA training configuration and run logs, an autoscaling serving deployment, CI/CD on every commit, production observability, an evaluation comparing base versus fine-tuned on a held-out set, a marketing landing page, a 10-slide pitch, a recorded demo, and a README documenting data sourcing, the contamination check, the scaling design, and the cost of the run.
Market signal — who wants thisDomain fine-tuning is a funded 2026 platform category: Together AI, Predibase (acquired by Rubrik in June 2025 for enterprise security depth), and Prem Studio compete on managed LoRA/QLoRA, and adapter-routing (one base model, many adapters per request) is now standard. The economics are compelling — a 7B model can be specialized on a single consumer GPU in an afternoon. Enterprises buy custom models that speak their technical language; investors fund fine-tuning platforms because every vertical AI product needs a model adapted to its own data.
Week 9
Align a model to human preference: reward modeling and PPO-based RLHF, and the simpler Direct Preference Optimization. What 'alignment training' changes and what it does not.
Bridges to Machine Learning — reinforcement learning and policy optimizationBuilds on: Parameter-Efficient Fine-Tuning: LoRA & QLoRA
Week 10
Compress a large teacher into a small, deployable student that keeps most of the capability. Distillation objectives, synthetic-data distillation, and honest accuracy accounting.
Bridges to Machine Learning — model compression and the teacher-student paradigmBuilds on: Preference Optimization: RLHF & DPO
Week 11
Benchmarks lie when misused. Build evaluation harnesses, control for contamination, measure on task-representative data, and report uncertainty instead of a single number.
Bridges to Statistics — sampling, confidence intervals, and experimental designBuilds on: Knowledge Distillation & Model Compression
Week 12
An enterprise mandate: a large model solves a task well but is too expensive to serve at volume. Distill its capability on that task into a small student model that can run cheaply, then ship the student as a launched product. Use the teacher to generate or label training data, train and align the student, and prove the student keeps most of the capability at a fraction of the cost. The deliverable is not a benchmark table — it is a directly deployable, hyperscalable product: the student served behind a real public API with a hyper-usable demo UI, autoscaling, CI/CD, observability, security, and full marketing (landing page, pitch, demo). Whatever time it takes — the deliverable is a model the business can actually afford to run and a product a buyer can try. Ship it as a real product.
Why it matters: Distillation is the standard route from an expensive frontier model to an economically viable product feature. A builder who can distill, align, and deploy a specialist student is directly deployable as a senior Model Engineer or Applied Scientist, a ₹1-crore-tier role because distillation work converts directly into serving-cost reduction at scale.
A publicly hosted product with a stable URL and a hyper-usable demo UI, plus a public repo and a published student model: the distillation data pipeline, the student training and preference-optimization configuration, an autoscaling serving deployment, CI/CD on every commit, production observability, a benchmark comparing teacher, student, and base on the target task, a cost-and-latency comparison, a marketing landing page, a 10-slide pitch, a recorded demo, and a README on the distillation method and scaling design.
Market signal — who wants thisDistillation drives the most-cited 2026 enterprise-AI economics: task-specific small models (0.6B-8B) match or beat frontier models at 10-100x lower inference cost, retaining 85-95% of capability. A $35K-$120K distillation project pays back in three weeks to three months against frontier inference bills, and startups like distil labs are funded purely to 'replace LLMs with custom small language models.' Investors back distillation because it is the clearest path from an expensive frontier model to a margin-positive product feature.