BLOG
Short notes on lessons from my projects — the kind of trade-offs that don't make it into a résumé bullet.
Written by me
Why I split my legal RAG into two models
~3 in 10 generated outputs invented realistic-looking filing dates. Here's why routing factual fields to extractive QA fixed it.
Read → Time Series · Multimodal MLWhen the simple model wins
My thesis benchmarked 10+ models for stock forecasting. The surprise: a standalone LSTM stayed competitive with fused multimodal models.
Read →AI THIS WEEK
Curated automatically from primary sources by a zero-dependency pipeline I built; each item links to its original source.
A weekly, auto-curated digest of AI developments — pulled straight from official lab blogs, arXiv, and GitHub. Every card links to its original source. Updated Jul 20, 2026.
A scorecard for the AI age
Sarah Friar, CFO of OpenAI, introduces a practical AI scorecard to measure ROI through useful work, cost per successful task, dependability, and return on compute.
Jul 17, 2026 · openai.com → Open Source · Hugging FaceFine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers
Read the full story at the source.
Jul 17, 2026 · huggingface.co → Research · MIT News — AIFollowing the questions where they lead
Assistant Professor Bailey Flanigan has arrived at complex computational methods for helping democracy thrive.
Jul 17, 2026 · news.mit.edu → Infrastructure · arXivPagedWeight: Efficient MoE LLM Serving with Dynamic Quality-Aware Weight Quantization
Mixture-of-Experts (MoE) is a popular class of large language models (LLMs), offering high efficiency and accuracy. However, in KV-cache-intensive serving scenarios, MoEs often exhibit a tension between the GPU memory…
Jul 17, 2026 · arxiv.org → Research · arXivA Blueprint for Equilibrium-Based Differentiable Continuous-Variable Thermodynamic Computing
To address the escalating energy and latency demands of machine-learning workloads, we introduce a blueprint for an energy-efficient and fast thermodynamic computing stack that leverages stochastic analog processes in…
Jul 17, 2026 · arxiv.org → Infrastructure · arXivCluster-Aware Matching via Laplacian Optimal Transport
In many applications of matching, the point clouds to be matched are not merely unstructured sets of points but rather samples from distributions with an intrinsic cluster structure. In such cases, as individual points…
Jul 17, 2026 · arxiv.org → Research · Google DeepMindOur approach to bioresilience
Google DeepMind and Isomorphic Labs are sharing our joint approach to bioresilience and AI models.
Jul 16, 2026 · deepmind.google → Product Update · Google AIConnect more of your apps to Search
You’ll be able to securely link and interact with your go-to services directly in AI Mode.
Jul 16, 2026 · blog.google → Product Update · MetaAlerting Parents if Teens Show Signs of Distress in Conversations With Meta AI
Parents using Instagram supervision tools will now be notified if their teen discusses suicide or self-harm with Meta AI. The post Alerting Parents if Teens Show Signs of Distress in Conversations With Meta AI appeared…
Jul 16, 2026 · about.fb.com → Model Release · AnthropicIntroducing Claude Tag
Read the full story at the source.
Jul 15, 2026 · anthropic.com → Model Release · Mistral AIYour Prompts and Skills need a system of record.
Studio provides a system of record for AI prompts and skills—versioned, owned, and traceable. Iterate fast, ship with control, and ensure consistent AI behavior.
Jul 9, 2026 · mistral.ai → Open Source · Berkeley AI ResearchIntelligence is Free, Now What? Data Systems for, of, and by Agents
... government of the people, by the people, for the people ... — Abraham Lincoln, Gettysburg Address (1863) The cost of AI is dropping rapidly. GPT-4-class capabilities cost roughly $30 per million tokens in early…
Jul 7, 2026 · bair.berkeley.edu →