Writing & Research

This is the 2025 finale: a practical reference architecture for running fleets of agents with governance—connectors you can trust, traces you can debug, evals you can ship, and humans you can hand off to.

Once agents can call tools, connectors become the new platform surface. This month is a playbook for adopting MCP at scale: patterns that work, versioning that doesn’t break customers, and governance that keeps the ecosystem sane.

Voice turns LLMs into real-time systems. This month is about building voice agents that meet latency budgets, degrade safely, and hand off to humans without losing context—or trust.

The EU AI Act turns “model choice” into a regulated interface. This month is a practical playbook: what GPAI providers must ship, what enterprises must demand, and how to build compliance into your agent platform without slowing delivery.

In 2025, the riskiest part of “agentic” systems isn’t the model — it’s the connectors. This month is a practical playbook for securing tools: least privilege, prompt-injection resistance, safe side effects, and auditability that holds up under incident response.

Multi-agent setups don’t fail because “agents are dumb.” They fail because we forgot distributed-systems basics: authority, contracts, budgets, and observability. This month is a practical architecture for scaling agents without scaling chaos.

Agents don’t “run in a loop” — they run across networks, vendors, and failures. This month is about the three patterns that make agent workflows survivable: durable intent (outbox), long-running transactions (sagas), and reconciliation (“eventually correct”).

Long context doesn’t kill RAG — it changes what’s cheap, what’s risky, and what needs architecture. This month is a practical guide to building “context-first” systems without shipping a cost bomb (or a data leak).

In 2025, “agents” stop being demos and start being products. This is the month you realize you don’t need a smarter model — you need a runtime: durable execution, safety gates, and traces you can debug.

Prohibited practices aren’t a legal footnote — they’re product constraints. This month is about turning “don’t do this” into guardrails you can deploy: policy gates, capability limits, audit trails, and incident-ready governance.

If your agent ships without tests, it’s not an agent — it’s a production incident with good marketing. This month is about turning “it seems fine” into eval gates you can run in CI.

Tool use was the warm-up. Computer-use agents can click, type, and navigate real UIs — which means mistakes become side effects. This article turns “agent can drive a screen” into an architecture you can defend: isolation, action gating, verification, and auditability.

Agents are becoming the new integration surface. This is how you go from bespoke tool wiring to an ecosystem: portable connectors, common protocols, and a practical standard called MCP.

Real-time agents aren’t “LLMs with voice.” They’re distributed systems with hard latency budgets. This month is about streaming UX, safe interruption (barge-in), and session state you can actually operate.

“Think step by step” isn’t an architecture. In production you need budgets, routing, and verifiers so the system knows when to go fast, when to slow down, and when to refuse.

The EU AI Act isn’t a PDF you “comply with” — it’s a set of control objectives you design into your product: evaluation, documentation, monitoring, and provable safety boundaries.

Tool use is where LLMs stop being “text generators” and start being integration surfaces. With open weights, reliability isn’t a given — you have to engineer it with contracts, sandboxes, and explicit capability boundaries.

RAG isn’t “add a vector DB and hope.” It’s a search-and-reasoning subsystem with contracts, metrics, and failure budgets — and you can only operate what you can evaluate.

Multimodal isn’t “a bigger prompt”. It’s a perception + reasoning + UX system with new contracts, new failure modes, and new latency/cost constraints. This month is about designing it so it behaves predictably.

Open weights shift your risk from vendor to you. This month is the playbook: evaluate like a product, treat licensing as architecture, and ship with guardrails that survive real users.

The moment you have multiple models with different strengths, “pick the best model” turns into system design. This month is a practical blueprint for routing, fallbacks, and budget-aware capability tiers—without turning your stack into an untestable mess.

“Stuff vs retrieve” is only half the battle. The operable part is a context assembler: a subsystem that selects, budgets, sanitizes, and logs exactly what the model sees—so you can debug, evaluate, and scale LLM features without vibes.

Bigger context windows tempt teams to paste everything. But long context is just a larger input buffer — not memory, not grounding, and not a plan. This month: how to budget context, decide “stuff vs retrieve,” and build a context assembler that stays fast, cheap, and safe.

Prompting feels like coding until it fails like statistics. This month I start treating LLMs as probabilistic components: define contracts, enforce schemas, and design failure budgets so your system survives outputs that are “plausible” but wrong.

Diffusion models made image generation possible. Midjourney made it feel addictive. The difference wasn’t just the model — it was the product loop: fast iteration, visible search, and UI as a steering system for probability.

DALL·E wasn’t “just a cool generator.” It turned language into a control surface for visual distributions — and forced product teams to treat image creation like a probabilistic runtime with safety, latency, and cost constraints.

Tool use is not “prompting better” — it’s turning an LLM into a controlled orchestrator of deterministic systems. This month is about the architecture, safety boundaries, and eval discipline that make agents shippable.

RAG isn’t “add a vector DB.” It’s a reliability architecture: define truth boundaries, build a testable retrieval pipeline, and evaluate groundedness like you mean it.

Hallucinations aren’t “the model lying”. They’re what happens when a probabilistic engine is forced to answer without enough grounding. This post is about designing products that stay truthful anyway.

ChatGPT isn’t “an LLM”. It’s a carefully designed product loop: context assembly, policy layers, tool orchestration, and observability wrapped around a probabilistic core.

RLHF is best understood as control engineering: a learned reward signal plus a constraint that keeps the model near its pretrained competence. Here’s how it works and how it fails.

Pretraining gives you a powerful text predictor. Instruction tuning turns it into something that behaves like a helpful tool. This post explains what instruction tuning changes, what it can’t change, and how to design products around the new failure modes.

Pretraining isn’t “learning facts.” It’s learning to compress a giant slice of the internet into a predictive machine. This post gives senior engineers the mental model: tokens, data mixtures, scaling, and why capabilities seem to ‘emerge’—plus the practical implications for cost, reliability, and product design.

RNNs made sequence learning feel like fighting gradients. Transformers made it feel like building systems: parallelism, short gradient paths, and a memory mechanism you can scale. This post explains attention as an engineering unlock—and what it implies for real software.

Before transformers, language models tried to compress entire histories into a single hidden state. This post explains why that was brittle: depth-in-time, vanishing/exploding gradients, and the engineering limits of “memory” — and why attention was inevitable.

LLMs aren’t “features” — they’re probabilistic runtime dependencies. This post gives the mental model, contracts, failure modes, and ship-ready checklists for building real products on top of them.

A production system isn’t “done” when it works — it’s done when it can fail, recover, evolve, and stay correct under pressure. This capstone stitches the 2021–2022 series into a reference architecture and a decision log you can defend.

Most teams “have on-call”. Fewer teams have resilience. This is a practical blueprint for designing systems, teams, and workflows that respond fast, recover safely, and learn without blame.

Reliability is non-negotiable, but “cost” is where architecture meets physics. This month is a practical playbook: how to model cost, allocate it, and design guardrails so your system scales without surprising invoices.

Most “data problems” are actually truth problems. This month is a practical mental model for product teams: where truth lives, how it moves, when to stream, when to batch, and how to keep analytics useful without corrupting production.

A practical mental model for choosing cloud primitives without ideology—based on responsibility boundaries, scaling, reliability, cost, and team operating capacity.

Performance isn’t a tuning phase — it’s an architecture property. This month I lay out an end-to-end mental model (browser → edge → app → data) and the practical playbook for improving both “fast on average” and “fast under load” without shipping fragile optimizations.

APIs don’t fail because they’re slow — they fail because they change. This month is about designing contracts you can evolve, enforcing compatibility automatically, and scaling teams without “everyone upgrade on Tuesday.”

Once your system crosses a process boundary, “a transaction” stops being a feature and becomes a strategy. This post is a practical mental model for distributed data: what to keep strongly consistent, what to make eventually consistent, and how to do it safely with outbox + sagas.

Security isn’t a checklist you add at the end — it’s a set of architectural constraints. This month is about threat modeling that fits real teams, and defaults that prevent whole classes of incidents.

Observability isn’t “add dashboards.” It’s designing feedback loops you can trust: signals that answer real questions, alerts tied to user pain, and tooling that helps you debug under pressure.

CI/CD isn’t a DevOps checkbox — it’s the architecture that makes change safe. This month is about test economics, pipeline design, and rollout strategies that turn “deploy” into a reversible decision.

Containers aren’t “how you deploy apps” — they’re how you make environments stop being a variable. This is the operational discipline: immutable artifacts, repeatable builds, and runtime contracts you can actually rely on.

Microservices aren’t a flex — they’re a tax. Modular monoliths aren’t “temporary” — they’re often the best architecture. Here’s the decision framework, the failure modes, and the migration path that doesn’t create a distributed mess.

Async work is where production gets honest. This month is a practical playbook for queues, retries, idempotency keys, and the patterns that keep “background jobs” from duplicating money or burning trust.

Caching is not “make it faster.” It’s a contract: what can be stale, for how long, for whom, and how you recover when it lies. This month is a practical architecture guide to caching layers that scale without corrupting truth.

Stop choosing databases by brand. Choose them by invariants, access patterns, and what “correct” means when the network is on fire.

REST isn’t “JSON over HTTP.” It’s a set of constraints that make interfaces boring, predictable, and resilient under change. This month is about designing resource boundaries and contracts that survive growth — and using versioning only when you’ve earned it.

Early on, any backend “works.” Then timeouts, GC pauses, cold starts, and operability show up. This is a practical mental model for choosing runtimes and frameworks based on constraints—not vibes.

The month I stop treating “frontend architecture” as component trivia and start treating it as a system: URLs as contracts, state as truth, forms as transactions, and a boring stack you can operate.

React is not a UI library — it’s a runtime for composing change. This post teaches the mental model that makes large React codebases predictable: component boundaries, hooks as architecture, and the real cost model of re-rendering (so you optimize the right thing).

Your data boundary is not “how we call the API.” It’s a coupling contract between teams, runtimes, and failure modes. This post gives you a practical decision framework for REST, GraphQL, and RPC-style boundaries (including tRPC) — with the tradeoffs that show up in production.

The browser is a constrained runtime with a scheduling problem: one main thread, many responsibilities, and users who notice missed frames. This post gives you the mental model to debug “random” UX failures as deterministic timing and contention issues.

HTTP isn’t just “how browsers talk to servers.” It’s a mature distributed-systems contract with semantics for caching, retries, concurrency, intermediaries, and evolution. If you design APIs without those semantics, production will teach you them anyway.

The web didn’t evolve because developers got bored. It evolved because latency, state, and economics kept forcing us to move responsibility between server, client, and edge. This post gives you the mental model and the checklist to choose the right rendering architecture in 2021+

2020 is when I stopped training agents and started building a trading system: environments, evaluation discipline, safety, and a live loop that survives outages. This is the postmortem — and the first result that actually held up in reality.

In late 2020 I stopped trusting hero backtests. I built a batch runner + a walk-forward evaluation harness, added eval gates, and discovered an uncomfortable truth: shorter training often wins.

After months of "all-in" agents with bull personalities, I rebuilt the environment to teach risk: stackable positions, time-awareness, and penalties that prevent reward-hacking.

In September 2020, I stop trying to be fast and start trying to be executable. The surprising result: in maker-style trading, fees aren’t a footnote — they can be the whole edge.

August 2020 - After the first live pain and the bull-personality problem, I stopped tuning "algorithms" and started tuning the network contract. Deep Silos beat flat MLPs, and the LSTM variant overfit fast.

After my first disappointing live runs, I stopped asking my agent to be clever and started forcing it to be safe: risk caps, timeouts, and “market-health” gates that kept the loop alive when the regime wasn’t.

Backtests looked amazing. Live PnL didn't. In June 2020 I ran the first real BitMEX live loop at tiny size and learned the most important lesson in trading ML: regime is the boss.

In May 2020, I stop hoping the bot is “fine” and start giving it explicit failure states — stale websockets, missing fills, rate-limits, and the kill switches that keep a live loop honest.

The moment RL stops being a notebook artifact: load a PPO policy, rebuild the live observation stream, and turn BitMEX into a runtime you can monitor and control.

In March 2020, I stopped treating reward like a number and started treating it like a contract—pay real costs, punish real risk, and don’t teach the agent to win a video game.

Training reward was lying to me. So I turned evaluation into a first-class system - chronological splits, deterministic runs, and walk-forward backtests that survive the next dataset.

In January 2020 I stop “predicting” and build a Gym environment that turns BitMEX microstructure features into actions, fills, and rewards — and makes every hidden assumption painfully visible.

I stopped pretending “a good predictor” was the same thing as “a tradable strategy” and designed a Gym-style environment contract that makes cheating obvious and failure modes measurable.

My first live alpha monitor was “working”… until BitMEX started replying 503 right when the model got excited. That’s when I learned availability is part of market microstructure.

I stop treating my alpha model like a notebook artifact and make it sit in the real BitMEX stream. The goal is not trading yet. It is seeing whether my features, normalization, and inference loop survive reality without quietly cheating.

My first serious attempt at making the model “see” microstructure features the way I designed them — grouped, compressed, and only then fused.

I train my first alpha predictor on BitMEX order-book features, and learn why ‘it trains’ is not the same as ‘it works’.

Before I train anything, I need a label that doesn’t smuggle the future into my dataset.

In June 2019 I stop treating feature scaling as “preprocessing” and start treating it as part of the production contract - same transforms, same stats, same order — or the live system lies.

If the order book is the battlefield, features are the sensors. This month I stop hand-waving and teach my pipeline to measure liquidity being added and removed - in a way I can deploy live.

In April 2019 I learned that the hardest part of trading ML isn’t the model — it’s the dataset contract. This month is about HDF5, integrity checks, and building rules that stop “good backtests” from lying.

In March 2019 I stop “talking about microstructure” and start collecting it. Websockets drop messages, clocks drift, and the only thing that matters is producing a snapshot I can trust.

If RL taught me “the state is the contract,” then trading is where that contract becomes painful. This month I map order book microstructure into concrete feature families my models can actually learn from.

In 2018 I learned RL inside clean Gym worlds. In 2019 I’m pointing that mindset at BitMEX — where the “environment” is a matching engine and the rewards come with slippage, fees, queue priority, and outages.

I ended 2018 in a weird place - less “reinforcement,” more “copying.” GAIL taught me that sometimes the fastest path to competence is to borrow behavior first — and ask questions later.

This month RL didn’t fail loudly. It failed quietly. Sparse rewards taught me the most brutal lesson yet - if nothing “good” happens, nothing gets learned — unless you rewrite what counts as experience.

After DDPG, I stopped thinking of RL instability as a surprise and started treating it like a design constraint. This month I learned why TRPO exists — and why PPO/PPO2 became the practical answer.

This month I left “toy” discrete actions and stepped into continuous control. DDPG looked like the perfect deal—until I learned what off-policy really costs.

After actor-critic finally felt “trainable,” I hit the next wall - RL doesn’t just fail—it fails in loops. This month is my map of the most common ways it breaks.

Policy gradients were honest but noisy. This month I added a critic—and for the first time in RL, training started to feel like something I could actually steer.

DQN taught me how fragile value learning can be. This month I tried something different - learn the policy directly. No Q-table. No value “crutch.” Just behavior, gradients, and a whole new set of failure modes.

This is the month I stopped reading about deep RL and started running it. DQN is simple enough to explain, hard enough to break, and perfect for learning Baselines like an engineer.

Tabular RL felt clean because you could see the truth in a table. The moment I replaced the table with a model, RL stopped being a neat algorithm and became a fragile system.

Tabular RL is the last time reinforcement learning feels clean. Values become literal tables, planning becomes explicit, and the “aha” moments arrive fast.

Bandits strip RL down to one tension—explore vs exploit—so I can stop confusing luck with learning and start building real intuition.

I’m starting 2018 by shifting from deep learning to reinforcement learning. The first lesson isn’t an algorithm — it’s that the data pipeline is the policy itself.

A year after finishing Andrew Ng’s classical ML course, I’m trying to separate enduring principles from deep learning-specific techniques—and decide where to go next.

After vanilla RNNs taught me why gradients collapse through time, LSTMs finally felt like an engineered solution - keep the memory path stable, and control it with gates.

RNNs looked elegant on paper. Training them exposed the same old enemy—vanishing/exploding gradients—just with “depth in time”.

Images were hard, but at least they were static. Sequences add “time”, shared weights, and state — and suddenly the assumptions I relied on in 2016 stop holding.

Convolution made CNNs "possible". Pooling and depth made them "useful" - invariance, hierarchies, and feature maps that start to look like learned vision primitives.

After months wrestling with training stability, I finally hit the next wall - fully-connected nets don’t “get” images. Convolutions felt like the first time architecture itself encoded domain knowledge.

In 2016, gradient descent felt like “the algorithm.” In deep learning, it’s just the beginning. This month I learned why momentum and careful learning rates are what make training "actually move".

After learning why gradients vanish, I discovered something even more unsettling - deep networks can fail before training even begins, simply because the starting scale is wrong.

April 2017 — I used to treat activation functions like a minor math choice. Then I saw how one change (ReLU) could decide whether a deep network learns at all.

I assumed “more layers” would just mean “more power.” Instead I discovered that depth introduces a new failure mode - gradients can disappear (or explode) long before the model learns anything useful.

Backprop stopped feeling like magic when I treated it like engineering - track shapes, follow the chain rule, and test gradients like you’d test any critical system.

After finishing Andrew Ng’s Machine Learning course, I start my deep learning journey by revisiting the perceptron and realizing neural networks begin with ideas I already understand.

I wrapped Andrew Ng’s ML course by building anomaly detection and a simple movie recommender—two patterns that show up everywhere in real systems.

Implement K-means clustering, compress an image to 16 colors, then use PCA to reduce dimensions and build eigenfaces.

I built my first spam classifier with SVMs—learning how C and sigma shape decision boundaries, and why linear SVMs scale surprisingly well.

The most practical assignment so far - diagnose bias vs variance using learning curves, tune lambda with validation curves, and build a repeatable “next action” playbook.

Implement backpropagation for a 2-layer neural network, verify gradients numerically, train on handwritten digits, and hit ~95% accuracy.

Build a multi-class digit classifier with one-vs-all logistic regression, then run a small neural network forward pass on the same dataset.

Overfitting is what happens when your model memorizes training data. Regularization is the practical tool that keeps it honest.

My first real classifier - predict admissions from exam scores with logistic regression, then learn why regularization matters on a non-linear dataset.

Notes — two ways to fit linear regression - iterative gradient descent vs one-shot normal equation. Same goal, different tradeoffs.

Notes from Andrew Ng’s ML course — extend linear regression to multiple features, learn feature scaling/mean normalization, and stop writing slow loops by vectorizing everything.

Notes from Andrew Ng’s ML course — plot the food-truck dataset, implement computeCost + gradientDescent in Octave, and build intuition with J(theta) visualizations.

In 2016, I’m documenting my journey through Andrew Ng’s Machine Learning course—building intuition, writing Octave code, and learning how to think in data.