Artifacts on Sina's Page

Transfer Learning Lab

Sun, 09 Aug 2026 00:00:00 +0000

A CPU-friendly set of experiments on transferability between tasks in neural networks.

The lab uses controlled MNIST and Fashion-MNIST variants with a shared CNN so each experiment can compare directly against a from-scratch baseline. It studies pretraining and fine-tuning, layer transferability, task-transfer matrices, negative transfer, catastrophic forgetting, multi-task synergy and interference, modular composition, and self-supervised pretext tasks.

A key goal is to make broad transfer-learning claims concrete enough to falsify. The executed notebooks include machine-checked claim assertions and experiments such as asymmetric task transfer, poisoned and frozen representations, replay against forgetting, gradient conflict between auxiliary tasks, and a pretrained encoder reused inside a compound two-digit architecture.

Frontend Lab

Thu, 06 Aug 2026 00:00:00 +0000

A teaching lab for programmers who already know how to code but want a principled mental model of the modern frontend stack.

The same small notes application is rebuilt eight times, with each technology introduced only after the limitation it solves becomes concrete: HTML → DOM scripting → HTTP/fetch → Vite + TypeScript → React → Next.js + Postgres → React Native → authenticated web/mobile clients with a shared typed API.

DSPy Lab

Tue, 04 Aug 2026 00:00:00 +0000

A hands-on course for understanding LLM programs as systems that can be specified, evaluated, and optimized rather than as collections of hand-written prompts.

The lab progresses from signatures and modules through evaluation, few-shot optimization, instruction optimization, feedback-driven optimization, agents, inference-time strategies, and observability. The recurring mental model is deliberately close to machine learning: define the computation graph, define the metric, then optimize the program against data.

The repository is organized as self-contained notebooks, including a capstone on HotPotQA and an under-the-LLM-layer deep dive.

Voice AI Lab

Tue, 21 Jul 2026 00:00:00 +0000

A broad practical curriculum for modern speech AI in both directions: speech → text and text → speech.

The recognition track covers Whisper-family models, open alternatives, cloud APIs, long-form transcription, translation, timestamps, forced alignment, diarization, VAD, and streaming. The generation track covers open TTS models, voice cloning, expressive and multi-speaker synthesis, cloud APIs, low-latency streaming, and realtime speech-to-speech systems.

The lab emphasizes comparative engineering: what model to reach for, what hardware and licensing constraints matter, how dependencies conflict, and what latency/quality trade-offs appear in practice. Heavier model families run in isolated environments so experiments remain reproducible rather than turning dependency management into the experiment.

Stat-ML Lab

Mon, 20 Jul 2026 00:00:00 +0000

A large hands-on curriculum for advanced applied statistics and machine learning, built around experiments where possible rather than slideware.

The notebooks cover statistical foundations, A/B testing, causal inference, tabular ML, forecasting, Bayesian methods, SQL and Spark, product analytics, finance and operations research, and ML in production.

Many modules use simulated data with known ground truth so assumptions and failure modes can be tested directly: peeking in experiments, ratio-metric inference, bad controls in causal models, leakage, calibration, drift, backtest overfitting, and more. The goal is to connect mathematical ideas to the judgment required in real systems and senior technical interviews.

LLM Text Generation Lab

Sun, 19 Jul 2026 00:00:00 +0000

An experiment-first lab for understanding how language models generate text and how inference-time computation changes results.

One track compares provider-exposed reasoning and sampling controls across hosted APIs. The other uses local Hugging Face models to inspect and modify the decoding loop directly: logits, greedy decoding, sampling, beam search, stopping criteria, and repetition controls.

The later experiments move beyond one-shot decoding into multiple candidates, self-consistency, verifier and critic loops, adaptive compute budgets, and model-agnostic search including Monte Carlo tree search. Measurements include task score, output length, latency, calls, cost, and provider-reported reasoning usage where available.

Agentic Systems Lab

Sat, 18 Jul 2026 00:00:00 +0000

An executable lab for learning how agentic systems are actually assembled while separating durable architectural ideas from fast-moving SDK syntax.

The core design principle is:

deterministic workflow spine + bounded agentic decisions + objective verification + human authority at consequential boundaries

The curriculum progresses from a local tool loop to routing, parallel fan-out/fan-in, evaluator–optimizer loops, independent verification, delegation and handoffs, persistent state, human gates, budgets, multi-agent orchestration, and protocol-level interoperability.

LLM Post-Training Lab

Tue, 07 Jul 2026 00:00:00 +0000

A controlled experimental setup for studying a question that is easy to state and surprisingly subtle in practice: what changes when preference optimization is done with imitation learning, offline preference objectives, or reinforcement learning?

The lab compares SFT, DPO, PPO, and GRPO from common starting policies across two tracks: helpfulness with learned reward models, and mathematical reasoning with verifiable rewards.

Rather than reporting a single score, the evaluation treats each checkpoint as a point on a frontier. It measures reward or task success against KL divergence from the reference policy, then adds diagnostics for capability retention, diversity and mode collapse, preference log-probabilities, and reward-model disagreement.