Blog
Production lessons and independent reference-design research across agents, retrieval, fine-tuning and MLOps, with diagrams and implementation detail.
By topic
9 posts
September 20, 2026 · 6 min read
Jev is a model that never generates text — it answers typed questions with calibrated probabilities. Where that fits in a LangGraph agent, with code, and where I would not use it.
September 5, 2026 · 8 min read
Every run of the pipeline becomes a nested Langfuse trace, and a judge model scores it with DeepEval. Neither adds a millisecond to the response, because scoring starts after the client already has the answer. The wiring: a per-request callback, a fire-and-forget task that cannot fail the request, a judge decoupled from the graph LLM, and an offline golden set for the metrics live traffic cannot compute.
September 4, 2026 · 7 min read
A skill is a folder with two files. SKILL.md becomes the system prompt, skill.yaml declares which sources the run may read, and the shared LangGraph pipeline runs unchanged on an isolated thread. Slash-invocable, like a command. Here is how a domain persona parametrizes four points of one graph without forking it, and the allowlist bug a code review found along the way.
6 posts
September 25, 2026 · 6 min read
Your agent re-derives the architecture from grep on every task, and guesses confidently when the fragments lie. Hand it the verified graph instead — over MCP, with file:line citations.
June 9, 2026 · 10 min read
How to store user preferences in Mem0, retrieve them by context tag before each response, and handle context switching when users shift between different tasks. Practical patterns with LangGraph integration.
May 18, 2026 · 4 min read
A comparative analysis of the leading AI agent frameworks in May 2026: Hermes' self-improvement, OpenClaw's massive ecosystem, and Paperclip's orchestration.
6 posts
September 3, 2026 · 7 min read
I built a RAG service for a regulated finance and legal practice and called it adaptive-rag. It isn't. There is no LLM router picking a path. Vector search always runs, web search is added only when a person flips a toggle, and four self-reflection gates with hard-capped loops decide whether the answer is good enough to ship. Here is why that design beat the classic one.
May 13, 2026 · 8 min read
How I built a real-time, hybrid voice AI: VAD, STT, and LLM run locally, while default Edge-TTS sends response text to Microsoft Edge's online TTS service.
April 21, 2026 · 6 min read
Learn how to design decoupled LLM systems that handle 10,000 requests per minute. Covers async queues, caching strategies, RAG integration, and the three biggest bottlenecks you'll face.
6 posts
June 15, 2026 · 7 min read
How I turned a prompt-driven ReAct agent over a 40+ tool MCP backend into a fine-tuned one — collecting multi-turn tool trajectories, building the SFT dataset, training Gemma 3 4B with QLoRA, and serving it back behind the agent loop. The case study that ties the whole fine-tuning series together.
June 14, 2026 · 7 min read
Full fine-tuning a 7B model needs ~112GB before you load a batch. LoRA trains ~1% of the weights; QLoRA squeezes the frozen base into 4 bits so a large model fits on a single card. Here's how both work — and the four pieces of QLoRA interviewers always probe.
June 13, 2026 · 6 min read
Supervised fine-tuning looks like 'teach the model the answer.' Mathematically that's not what happens. It's the same next-token prediction as pretraining — and two quiet details, response masking and the chat template, decide whether your fine-tune works or silently rots.
September 25, 2026 · 7 min read
A clean-looking PR rerouted around a load-bearing layer and the diff didn't show it. The lockfile design, --diff, --drift-base, and CI setup that came out of that review.
September 25, 2026 · 6 min read
Install svarupa, scan a repo, and read the evidence-backed report: artifact contents, the viewer, the query CLI, and the MCP server — anchored on two live demo reports.
September 25, 2026 · 6 min read
Your agent re-derives the architecture from grep on every task, and guesses confidently when the fragments lie. Hand it the verified graph instead — over MCP, with file:line citations.
September 25, 2026 · 7 min read
I published a CLI that renders architecture diagrams where every node and edge carries file:line evidence — and stays silent where it has none. This is the thesis behind it.
September 20, 2026 · 6 min read
Jev is a model that never generates text — it answers typed questions with calibrated probabilities. Where that fits in a LangGraph agent, with code, and where I would not use it.
September 17, 2026 · 9 min read
A decision framework for applying governed transaction controls across procurement, travel, paid APIs, recurring operations, and service purchasing.
September 17, 2026 · 8 min read
A practical model for turning an approval click into typed, transaction-bound authority that deterministic controls can verify.
September 17, 2026 · 8 min read
Why ambiguous transaction outcomes must pause blind retries and move through evidence-led reconciliation.
September 17, 2026 · 8 min read
How to consume approval, reserve constraints, issue scoped authority, and hand off evidence without pretending an external side effect is atomic.
September 17, 2026 · 8 min read
A target design for workload identity, credential brokering, short-lived grants, audience restriction, and isolated execution.
September 17, 2026 · 10 min read
A fictional purchase traced through identity, policy, approval, scoped execution, uncertain outcomes, reconciliation, and final evidence.
September 17, 2026 · 9 min read
How deterministic spend policy combines action limits, budgets, supplier rules, validity, velocity, risk, and separation of duties.
September 17, 2026 · 8 min read
An illustrative procurement architecture that separates agent planning, spend policy, exact approval, isolated purchasing, and receipt evidence.
September 5, 2026 · 8 min read
Every run of the pipeline becomes a nested Langfuse trace, and a judge model scores it with DeepEval. Neither adds a millisecond to the response, because scoring starts after the client already has the answer. The wiring: a per-request callback, a fire-and-forget task that cannot fail the request, a judge decoupled from the graph LLM, and an offline golden set for the metrics live traffic cannot compute.
September 4, 2026 · 7 min read
A skill is a folder with two files. SKILL.md becomes the system prompt, skill.yaml declares which sources the run may read, and the shared LangGraph pipeline runs unchanged on an isolated thread. Slash-invocable, like a command. Here is how a domain persona parametrizes four points of one graph without forking it, and the allowlist bug a code review found along the way.
September 3, 2026 · 7 min read
I built a RAG service for a regulated finance and legal practice and called it adaptive-rag. It isn't. There is no LLM router picking a path. Vector search always runs, web search is added only when a person flips a toggle, and four self-reflection gates with hard-capped loops decide whether the answer is good enough to ship. Here is why that design beat the classic one.
June 15, 2026 · 7 min read
How I turned a prompt-driven ReAct agent over a 40+ tool MCP backend into a fine-tuned one — collecting multi-turn tool trajectories, building the SFT dataset, training Gemma 3 4B with QLoRA, and serving it back behind the agent loop. The case study that ties the whole fine-tuning series together.
June 14, 2026 · 7 min read
Full fine-tuning a 7B model needs ~112GB before you load a batch. LoRA trains ~1% of the weights; QLoRA squeezes the frozen base into 4 bits so a large model fits on a single card. Here's how both work — and the four pieces of QLoRA interviewers always probe.
June 13, 2026 · 6 min read
Supervised fine-tuning looks like 'teach the model the answer.' Mathematically that's not what happens. It's the same next-token prediction as pretraining — and two quiet details, response masking and the chat template, decide whether your fine-tune works or silently rots.
June 12, 2026 · 5 min read
Most people learn q_proj, k_proj, v_proj, o_proj and think that's the whole transformer. It isn't. A block has two machines — attention and the MLP — and knowing which does what is the difference between a LoRA config that learns your domain and one that doesn't.
June 11, 2026 · 6 min read
LoRA, SFT, QLoRA, DPO, PPO, GRPO — they all blur together until you see they live on two independent axes. One decides how you touch the weights, the other decides what signal you train on. The map I use before every fine-tuning project.
June 9, 2026 · 10 min read
How to store user preferences in Mem0, retrieve them by context tag before each response, and handle context switching when users shift between different tasks. Practical patterns with LangGraph integration.
June 2, 2026 · 8 min read
How I built a verification pipeline that catches hallucinated citations before they reach production. NLI cross-encoder, Claude judge, hard-fail gates, and the economics of getting it right at $0.05 per query.
May 18, 2026 · 4 min read
A comparative analysis of the leading AI agent frameworks in May 2026: Hermes' self-improvement, OpenClaw's massive ecosystem, and Paperclip's orchestration.
May 13, 2026 · 8 min read
How I built a real-time, hybrid voice AI: VAD, STT, and LLM run locally, while default Edge-TTS sends response text to Microsoft Edge's online TTS service.
April 25, 2026 · 14 min read
Stop guessing which LLM technique to use. Data-driven decision framework with ROI analysis, cost comparisons, and real production case studies from Document Extraction Pipeline and DeepAgent.
April 24, 2026 · 14 min read
Stop eyeballing LLM outputs. Learn production-grade evaluation: LLM-as-judge, deterministic checks, RAG faithfulness, and A/B testing frameworks that actually work.
April 23, 2026 · 11 min read
Stop letting context windows bankrupt your AI budget. Learn sliding window, summarization, and vectorized memory strategies with real cost savings from production deployments.
April 22, 2026 · 8 min read
Building on Document Extraction Pipeline's Celery/Redis foundation, learn to extend async patterns to LLM-specific resilience: circuit breakers, multi-provider fallbacks, and token bucket rate limiting.
April 21, 2026 · 6 min read
Learn how to design decoupled LLM systems that handle 10,000 requests per minute. Covers async queues, caching strategies, RAG integration, and the three biggest bottlenecks you'll face.
April 13, 2026 · 14 min read
Install OpenClaw, wire Anthropic/OpenAI/Google or a local Ollama model, control it from Telegram and Discord, extend it with skills.sh, and turn it into a business gateway for lead qualification, customer support, and agent-ecosystem management.
April 13, 2026 · 7 min read
How I built a production AI agent API using LangGraph, Ollama, and FastAPI with real-time SSE streaming, Redis rate limiting, and multi-turn session memory.
April 9, 2026 · 4 min read
A practical guide to building and deploying LLM-powered agents using LangChain and FastAPI, with lessons from real-world implementations.
April 9, 2026 · 7 min read
How I built a high-performance document extraction system using FastAPI, Celery, MinerU, and LLMs. Learn the architecture, challenges, and lessons from deploying AI-powered document processing at scale.
Tell me about it. I reply within one working day with a first take and no sales pitch.
