My journey in the AI and Agentic universe
I expected a tiny RL experiment to pick a winner. It turned into a build of five post-training paths, two inference engines, and a routed OpenAI-compatible provider. The result was flat across three seeds, but the learning was in the system required to make that fact impossible to hide.
A reference architecture for giving autonomous agents scoped authority and KMS-protected credentials without placing plaintext in etcd or durable workflow history. AgentShield joins OIDC, DPoP, workload identity, policy, and Temporal into one delegated-trust boundary.
We already know how to handle unpredictable actors with access to company systems. A reference identity gateway that binds Auth0 M2M governance, real-time admin gating, and Langfuse execution traces into a single auditable proof.
An execution agent should not grade itself. A small runtime assurance plane separates execution, hard safeguards, contextual judgment, and deterministic intervention—then makes the reason for every decision visible to a human.
On four Arm cores, BEAM coordination took 74 µs p99 against 2.206 ms p99 of selected-expert compute. This is the measured path from logical sparsity to bounded physical memory — including the prefetch and broker-delivery ideas that failed their gates.
I put an AI agent inside commercial loan underwriting and watched it approve a request it should have denied — fluently, quietly, past every safeguard. This is the system I built around it: an evaluation suite it cannot see, a two-tier correction ladder it cannot own, and the day it shipped a sandbox-verified code fix that I merged like any engineer's PR. 75% to 100%, under fifty cents, human authority intact at every step.
A 0.5B model handed a broken Rust file fixed nothing — zero out of forty-five. This is the story of what I put around it: a verified repair loop where the compiler is the judge, a failure-driven curriculum, and the night a nano model beat the flash-class model that taught it — 72% on frozen out-of-distribution repair, on a CPU, for under six dollars.
The confession after Phase 4: a gateway that works but leaks API keys in error logs is an incident waiting to happen. Five milestones of proving a five-package LLM platform will not leak, crash under dependency failure, or run unbounded bills — secret scrubbing at the logger boundary, PII guardrails, chaos engineering, and Stripe billing.
What happens after the prompt works: the outage you didn't plan for, the bill nobody can explain, and the question “why did it say that?”. Rebuilding the layer under every AI app into five production packages: a typed gateway, a hybrid RAG pipeline, an agent runtime, a control plane, and a shared contract library.
A blind debugging system powered by GPT-5.6 Sol and three Solari primitives. It reproduced 25 failures, repaired all four injected faults, passed 53 tests, and taught me why an agent must never be allowed to change the proof of its own success.
A living, visual encyclopedia of AI engineering: CUDA kernels, inference pipelines, and the systems behind modern machine intelligence, explained through diagrams rather than prose. I will keep building and adding entries over the coming months, so check back as the atlas grows.
Every agent framework ships with a fixed toolbox. FORGE ships with a forge: when it hits a capability gap, it writes a new tool, tests it in a sandbox, and registers it forever. Then came the experiment every skeptic would demand, a warm vs cold ablation on a live LLM. Does the growing library actually make it measurably better?
Stratum began as a confession: hundreds of files of simulated intelligence, beautiful dashboards, and not one real file changed. This is the story of tearing that down and building an execution-first agentic runtime where every transition emits a durable event, a human approval gate sits structurally between plan and side effect, and any run can be replayed from its receipts — no AI calls, no side effects.
I rebuilt the mechanisms behind DeepSeek-V4's million-token architecture at 1/2,500,000th scale: 13 experiments, 65 tests, one 629K-parameter integrated chain. The lesson was economic, not architectural: every mechanism is an exchange rate between memory and recall, bandwidth and silence, precision and adaptation. This is the full ledger.
An in-depth look at Aegis CLM, a zero-trust, multi-agent contract lifecycle platform engineered with bilingual PII redaction, corporate RAG playbooks, and Human-in-the-Loop interlocks to turn weeks of contract friction into governed commercial velocity.
I began with K3's headline scale and a component-by-component learning map. Both hid the real story: how memory, sparse capacity, depth, post-training, infrastructure, and evaluation constrain one another. This is the architectural journey that turned a paper summary into a complete learning system.
Exploring pm-graph, a platform designed to accelerate the product development lifecycle (PDLC) by treating Product Management as an Agentic Service (PDaaS) and orchestrating unified alignment across specs, backlogs, and A2A code synthesis.
Exploring a cohesive, production-grade 7-platform transformation stack. By treating agentic AI as robust infrastructure managed via a central control plane (AgentOps), we showcase an elegant blueprint for scaling autonomous agents across systems, data, and processes.
Can independent AI and human participants coordinate through durable events without a central orchestrator? The surprising lifecycle findings point to what comes next.
An in-depth look at building PWC-Curator, a self-hosted workspace that parses double-column research paper PDFs, extracts benchmarks and repository links, and performs hybrid lexical-vector searches with a dynamic slider.
Exploring the engineering and design patterns behind ARI-bench, an adversarial safety evaluation benchmark implementing Google's ASAT roadmap and DeepMind's TRAIT&R taxonomies across 4 testing tracks.