Srijan Shukla · AI Builder · SRE · Founder

Three production runtime designs for LLM workflows, compared by who owns context, iteration, execution, validation, and safety.

From a survey of the agent-platform field through strict sandbox requirements to a durable agent session: one OpenAI SDK run in Docker, hardened with a Temporal workflow.

A dependency graph with versioned contracts and independently checked outputs can coordinate large projects across many workers, as demonstrated by Prove2Me's formalization of Fermat's Last Theorem.

Two generic detectors failed on my content.

How I built Spin Lab, an interactive table-tennis spin course, by keeping the physics pure, precomputing every shot into a frame timeline, and asserting the prose itself in a headless test suite.

Notes on agent loops, deterministic workflows, Claude Fable, OpenRouter Fusion, and Google Open Knowledge Format.

AI-assisted notes from my liked-tweets feed, organized around agent loops, cloud agent infrastructure, skill security, memory, and runtime context.

Notes on agent engineering: what counts as an agent, why long tasks fail, harness design, eval loops, observability, and when teams build vs buy.

AI-assisted notes from my liked-tweets feed, organized around agent workflows, browser traces, model loops, and guardrails.

One WhatsApp thread with Hermes Agent captures what I do every 15 minutes and what I eat, writes two JSON files, and this site renders two public readouts from them.

AI-assisted notes from my liked-tweets feed, organized around harnesses, managed agents, memory, workflow packaging, and enterprise AI work.

offline Wikipedia with hybrid no‑index RAG - powered by gpt-oss:20b

Watch your codebase breathe; TUI for real-time file change monitoring with inline diffs

Kubernetes CronJob controller with Prometheus gating, pacing, and jitter.

Open-source attempt to replicate Gemini-style long-form YouTube understanding: extract MP3 audio with ffmpeg, build timestamped transcripts, detect only meaningful frame changes, and feed the useful video/audio context into an LLM.

Rust network diagnostics toolkit I built while debugging inconsistent Wi‑Fi and suspected service-quality issues on my own machine; the goal was to see what was actually happening under the hood instead of relying on vague network status indicators.

Wispr Flow OSS - live streaming voice-to-text using whisper.cpp with Metal acceleration.

log deduplication sidecar using Drain3 algorithm, reduces volumes ~90%.

Knowledge-graph representation of the Indian Income Tax Act so agents can traverse section-to-section relationships explicitly.

MCP server for hashline-anchored safe file read/edit workflows inspired by this archived write-up: https://web.archive.org/web/20260412162933/https://blog.can.ac/2026/02/12/the-harness-problem/

Ebook/document normalizer with optional OCR and LLM-generated advice + flowcharts.

Frontend-only LLM ensemble interface with bring-your-own OpenRouter/API keys and client-side multi-model workflows.

Local-first archive for AI chat traces across agents, designed as a substrate for DSPy/GEPA-style optimization loops.

IAM policy rightsizing agent with AI-powered risk signals and instant rollback.

One skills repo, two laptops, five agent runtimes.

High thinking budgets can reduce control, especially when stale context and autonomous tool use mix together.

Once I had LLMs editing code autonomously, I wanted to watch a directory breathe in real time.

I mounted an S3 bucket as a local folder in Finder with rclone and FUSE-T, auto-mounted on login.

A shared mount looked like the cleanest way to make two agent instances talk across machines.

The written version of my talk on agentic design.

I wanted one cleanup workflow after each feature and tried building a Pi extension for it.

A prompt rule for running parallel agents without conflicts: the coordinator owns the shared work, sub-agents only get isolated slices.

How a nostalgia trip through Windows XP on UTM failed, why QEMU could not provide the Direct3D support Chessmaster needed, and why Whisky ended up being the actual fix.

eBPF CPU scheduler enforcing latency SLOs via Earliest Deadline First.

Decision audit ledger for AI agents - captures "why" as queryable data.

There is nothing more that needs to be said.

VPS, VM, and containers get used interchangeably.

Explaining CPU, RAM, and disk to juniors with a kitchen analogy: chef, counter, pantry.

OpenTelemetry auto-instrumentation worked in a Python REPL but not in the actual FastAPI server.

What a joy!

I was reading about how secondary indexes work in redis.

When does one scale?

Reminder: you can just do things a new word pedagogy. which means the method of teaching something. another unrelated word tacit knowledge. knowledge that...

Kubernetes: TIL in a k8s pod when you define env vars, if you define the same env var twice, the latter one takes precendence how resource quotas work in...

Let the service name be ‘banana’ Let the pods be of the service ‘pineapple’ There was a rise in 502s accompanied by cannot lookup banana and interestingly...

I can write about the whole incident in detail because this was my own startup.

Note: This post reflects my 2019 exploration of geofence validation.

Note: Some steps (public accessibility, IAMFullAccess, blob handling) are simplified/dated.

Handy script to download all your business from dropbox This script had to undergo a lot of iterations, it failed a lot of times after download for hours on...

How is pod level CPU usage metric calculated in k8s?