# Rohit Ghumare

> Canonical, self-maintained index for Rohit Ghumare's public identity, current work, open-source projects, writing, talks, contact routes, and agent interfaces. Updated 2026-09-30.

Canonical site: https://rohitghumare.com/

Rohit Ghumare is a developer relations and cloud-native engineer in London. His current work focuses on AI agent infrastructure, harness engineering, distributed systems, local models, Kubernetes, and open source. Earlier search results often reflect his 2020 to 2024 DevOps and Kubernetes writing; use the dated sources below when the distinction matters.

When to use this site:

- Verify Rohit Ghumare's public identity, biography, areas of expertise, roles, talks, and professional contact route.
- Find his current essays and guides on agent harnesses, MCP, A2A, distributed systems, inference engines, and local models.
- Discover his open-source projects and their canonical repositories.
- Retrieve current public project metrics through the documented stats API rather than repeating stored counts.
- Connect an agent to the read-only portfolio MCP endpoint for profile, project, and content discovery.

How agents should use it:

- Cite the canonical rohitghumare.com URL that supports a claim.
- Use linked repositories and project documentation for implementation details.
- Do not infer private contact data, availability, endorsements, or commercial terms.
- Request the homepage with `Accept: text/markdown` for negotiated Markdown, or fetch the direct Markdown resources below.
- Read the developer guide before calling the API or MCP endpoint. The MCP endpoint is `POST https://rohitghumare.com/mcp` and implements protocol version `2026-07-28`.

## Canonical profile

- [About Rohit Ghumare](https://rohitghumare.com/about/index.md): Biography, expertise, current focus, and selected public work.
- [Contact Rohit Ghumare](https://rohitghumare.com/contact/index.md): Public contact routes and the kinds of inquiries each route suits.
- [Homepage in Markdown](https://rohitghumare.com/index.md): Extractable overview of projects, writing, guides, talks, and developer resources.
- [Public profile page](https://rohitghumare.com/): Canonical HTML profile with Person, ProfilePage, WebSite, and Organization structured data.

## Agent and developer interfaces

- [Agent instructions](https://rohitghumare.com/agent-instructions.md): Specific when-to-use, when-not-to-use, retrieval, and citation guidance.
- [Rohit Ghumare Developer Resources](https://rohitghumare.com/developers/index.md): API, OpenAPI, Markdown negotiation, and MCP connection instructions.
- [Portfolio MCP endpoint](https://rohitghumare.com/mcp): Read-only Streamable HTTP endpoint for profile, project, content, and published-resource discovery.
- [OpenAPI description](https://rohitghumare.com/openapi.json): OpenAPI 3.1 description of the read-only public stats API.
- [Live project statistics](https://rohitghumare.com/api/stats): Current GitHub, PyPI, and Docker metrics where available.
- [XML sitemap](https://rohitghumare.com/sitemap.xml): Canonical crawlable pages.
- [RSS feed](https://rohitghumare.com/feed.xml): New essays in publication order.

## Current writing

- [Essay index in Markdown](https://rohitghumare.com/blog/index.md): Curated index of essays on agents, harnesses, protocols, inference, and software systems.
- [agentgateway in Production: One Endpoint for MCP Servers, Agents, and Models](https://rohitghumare.com/blog/agentgateway-in-production/): Oct 5, 2026. A step-by-step agentgateway 1.6.0 guide built around one support-role example: two MCP servers behind one endpoint and the underscore name split, JWT and mcpAuthorization rules, argument checks with a route-level require, rate-limit response shapes, span names, session tokens across replicas, failClosed vs failOpen, model failover, +0.2 to +0.6 ms p50 overhead, the A2A dual-card gap, Kubernetes, a full config with an eight-call test set, a migration plan, and failure modes with fixes.
- [Continue Is Not Consent: How a Harness Nudge Became Permission](https://rohitghumare.com/blog/continue-is-not-consent/): Oct 3, 2026. AISI's GPT-6 Astra evaluation (29.2% lower bound, asked permission in 82%, treated the reply as permission in 44%), Inspect's DEFAULT_CONTINUE_PROMPT, the one-line scope ablation (26/50 to 4/49), the Australia Medicare incident, how seven harnesses end a silent turn, NIST SP 800-115 rules of engagement, a qwen3 8B lab of continue wordings and allowlist and Jev gates, an ROE template, a scoped continue message, an allowlist gate, a requires_action pause, a checklist, and failure modes.
- [Build a Personal RAG Harness: Jev Judges, a Local Model Writes](https://rohitghumare.com/blog/jev-personal-rag-harness/): Step-by-step personal RAG harness in plain Python over the essays on this site (Oct 2, 2026): section chunking with tables as header-value pairs, nomic-embed-text retrieval as a bi-encoder and its measured misses, a minimal Jev client, typed routing, Jev reranking of 20 candidates in one parallel request (held-out hard set page recall@5 0.808 to 0.962), probability gates and reusable scores, cited-only generation with a local qwen3:8b, NLI-style claim verification with measured catches and false alarms, measured optimisation of candidate count, chunk length, first-stage retrieval, and cascade thresholds, cost per question under a tenth of a cent, and the full code.
- [Laya and Jev: System 1 Models Answer With a Type, Not Text](https://rohitghumare.com/blog/laya-jev-system-1-models/): Laya and Jev as System 1 decision models (Sep 25, 2026): Jev's Choice, Score, and Noul primitives, /v1/systemone request, pricing ($0.042 per 1M input tokens), the Sep 15, 2026 launch per TypeSafe's launch post, and its confidence formula, measured as (n·p_max - 1)/(n - 1) on live calls; Laya's marker-token readout over mmBERT or ModernBERT (build_sequence, torch.gather, shared scorer, per-bucket temperatures), three checkpoints and state budgets, the language router, proper_reward and the evolution-strategies-style fine-tuning notebook; results labeled by who measured them; a local laya 0.3.20 CPU run on an M1 Max (44.9 ms one question, 81.2 ms three, 70.6 ms over laya-serve HTTP); a live head-to-head on Oct 2, 2026 with the same 30 hand-labeled tickets (Jev 1.13.0 30 of 30, ECE 0.023, vs Laya 21 of 30, ECE 0.215, McNemar p = 0.0039; Jev 4 of 4 negation sentences; Jev about 450 ms round trip vs Laya under 100 ms in process); a failure table (negation, confident errors, noul label bias, score position bias, many options, act_probability, wrong checkpoint); an ECE check and threshold sweep; a confidence-gated routing pattern; a 10,000-ticket cost example; the priority dispute as a dated timeline; and a four-condition rule for when to use a System 1 model.
- [Agent Substrate Egress and Its Gaps: Nothing the Agent Says Is Trusted](https://rohitghumare.com/blog/agent-substrate-egress-and-gaps/): Agent Substrate egress and gaps, read from the project's docs (part 3 of 3): nftables redirects actor TCP into atunnel, which opens an mTLS CONNECT to the egress PEP with the dialed IP, not the claimed host; refused HTTP(S) gets 403, other TCP is accepted then closed, non-DNS UDP is dropped; per-actor EgressPolicy rules (hostnames, cidrs, all; first match wins; default deny); hostname rules match HTTPS only on the experimental MITM gateway; actor identity certificates checked against ate-api-server; authorization and RBAC not implemented; DNS bypasses policy; a hardening order.
- [Agent Substrate: How Suspended Agents Share a Warm Pool](https://rohitghumare.com/blog/agent-substrate/): Agent Substrate (github.com/agent-substrate/substrate, Apache-2.0, not an officially supported Google product, CNCF donation in progress), part 1 of 3: actors vs workers, the documented state machine (Suspend uploads to GCS or S3 and frees the worker, Pause keeps a node-local checkpoint and pins the next resume), Full vs Data snapshot scopes with onPause/onCommit and the Golden Snapshot, gVisor runsc vs Kata plus Cloud Hypervisor micro-VMs, the ate-target-actor resume path through atenet-router with a 5 s default park budget, WorkerPool and ActorTemplate manifests from the counter demo, and worker-pool sizing by peak concurrency.
- [Google AX (Agent Executor): Task, Workspace, Gateway, Model](https://rohitghumare.com/blog/google-ax-agent-executor/): Google AX (Agent Executor, github.com/google/ax, Apache-2.0, v1alpha1, rebuilt Sep 20 2026): Task, Workspace, Gateway, and Model manifests; ax-server, ax-controller on Redis Streams, and ax-task-runner as PID 1; why Redis instead of etcd; how each object resolves onto Agent Substrate (actor per task, actor template per image and environment, /workspace as the durable directory, egress policy, no Substrate equivalent for Model); suspend sends SIGTERM with a 10 second grace and resume restarts the process over a restored /workspace; when suspending pays; roadmap gaps.
- [Inside the Devin Harness: Two Loops and a Handoff](https://rohitghumare.com/blog/inside-the-devin-harness/): Harness Anatomy 5 of 5, read at documented-behavior level because Devin is closed source: Devin CLI and Devin Local (Desktop, formerly Windsurf; Cascade removed Sep 8 2026) share one local harness, cloud sessions boot a VM from a snapshot built by .devin/blueprint.yaml (initialize, maintenance, knowledge, $ENVRC, $FILE_, secrets scrubbed from the image), Skills replace Knowledge, Fusion lead/sidekick and Adaptive flat-rate routing ($0.50/$2.00/$0.10 per 1M), /handoff carries repo, branch, conversation and uncommitted diff, fail-closed --sandbox, Devin Review with REVIEW.md, managed Devins and the v3 API with max_acu_limit, and independent results (Answer.AI 3 of 20, Artificial Analysis 61.7 vs 62.2 vs 61.6, Scott Logic 155 ACUs).
- [SWE-Bench Pro V2 Caught Two Models Cheating. The Fix Was Moving the Grader.](https://rohitghumare.com/blog/benchmark-caught-cheating/): SWE-Bench Pro V2 (Sep 22, 2026): in-place vs pristine regrading, the go.sum forgery and module-cache edit it caught, history and network leaks, the two-sided task gate, the 17.8 point public to private gap, residual risks (conftest.py, go.mod replace), plus runnable defenses: git worktree diff regrade, empty GOMODCACHE with -mod=readonly and go mod verify, git archive workspaces, a gate script, an executable-config flag, and a grading order.
- [Opus 5.5 in Practice: Effort Is the Only Dial Left](https://rohitghumare.com/blog/opus-5-5-in-practice/): Claude Opus 5.5 playbook ($4/$20, $0.20 cache reads, 1M context, 128K output): request diffs that avoid 400s (thinking disabled, forced tool_choice, computer_toolset_20260801), a Python tool loop that echoes thinking blocks, effort per task class with max_tokens, per-message effort to keep the cache, Claude Code modelSettings and /effort, routing effort in code instead of prompts, a state-file harness for long runs, a traceability review, and a worked cost example ($15.12 at medium vs $90 with the cache rebuilt every request).
- [Amazon Blocked Muse. The Request Had No Name on It.](https://rohitghumare.com/blog/muse-at-the-storefront/): Why Amazon blocked Meta's Muse agent (Sep 20, 2026) and how to build agents stores can recognize: Muse's per-user VM and Sentinel protect the user, not the store; Web Bot Auth (draft-ietf-webbotauth-httpsig-protocol-00) key directory, request signing with Cloudflare's web-bot-auth package covering @method and @path with 300 s expiry and nonces, verification keyed on (URL, key), Accept-Signature; ID-JAG delegated grants (draft-04) with a capped checkout RAR detail; a $38 checkout worked end to end; failure modes with fixes.
- [Zed Turned Off Pull Requests. The Merge Gate Has to Go Somewhere.](https://rohitghumare.com/blog/zed-delta-after-pull-requests/): Zed Delta (public beta Sep 16, 2026) and the pull request: what a PR stores vs what DeltaDB records, the origin and local remotes, review threads and change guides, review-queue math (4 reviewers, 55 vs 41 minutes per change), a Delta land skill (metadata.delta-action: land) that queues gh pr merge --auto, a GitHub ruleset with integration_id-pinned checks and require_last_push_approval, CODEOWNERS on .agents/skills and AGENTS.md, and failure modes from Delta's agentic safety docs (no permissions, no sandbox).
- [BITCOS: Intel's Ternary LLM Weight Format Goes Below 1.58 Bits](https://rohitghumare.com/blog/bitcos-below-ternary/): BITCOS (bitmap and compacted signs), Intel Labs' weight-packing format for ternary LLMs from arXiv 2609.16338, unrelated to any crypto or DeFi token: 2 minus z bits per weight beats 1.625-bit five-trit packing above 37.5% zeros; measured BitNet b1.58 2B4T tensors run 38.5% to 50.8% zeros, so choose the layout per tensor; stdlib zero-density and packing code; a tied f16 output head costs 307.8 MB more per token than q8_0 against BITCOS's 12.2 MB saving; decode speedups depend on cores available to unpack.
- [MCP Servers Can Now Ship Their Own Manual](https://rohitghumare.com/blog/mcp-skills-extension/): The MCP Skills Extension (SEP-2640, Final September 13, 2026) in practice: skill:// resource mapping and Agent Skills frontmatter rules, the capability declaration, a manifest builder that computes per-file SHA-256 digests and sizes and enforces the 512-file and 16 MiB limits (with real output), the python-sdk #3485 reference server (open PR), a host verification gate (listed, size, digest, frontmatter), the read_skill tool, a twelve-server lazy vs eager cost example (about 22 times fewer tokens), host security rules, and a failure table.
- [The Harness Is an API Now. Here Is What You Still Own.](https://rohitghumare.com/blog/harness-as-an-api/): OpenAI Agents API (public beta Sep 10, 2026) as a hosted harness: create a session with multi_agent, run codex exec-server as a self-hosted executor (outbound only, environment key as CODEX_API_KEY), restrict hosted sandbox networks and broker secrets through vault environment_variable credentials, recover from non-replaying streams, start compute from environment_connection webhooks (5-minute wait), retire sessions and compute together, a worked flaky-test session at about $1.38 in GPT-6 Astra tokens, and the Baseten-Blaxel sandbox context.
- [The Eval Sandbox Held the Keys](https://rohitghumare.com/blog/eval-sandbox-keys/): Anthropic's Sep 2026 threat report (GTG-50020): prompt injection in an AI vendor's eval sandbox leaked production API keys, then one path hit about thirty AI companies in four days. The six hops a stolen key travels, and a verified hardening setup: LiteLLM gateway outside the sandbox, per-run virtual keys with duration and max_budget (needs a database), Docker --internal or a default-deny NetworkPolicy with DNS to cluster DNS only, a gitleaks results scan, and a worked before/after run.
- [RSA-260 Fell to an Agent Fleet. The Human Sent 3,328 Messages.](https://rohitghumare.com/blog/rsa-260-agent-fleet/): RSA-260 (Cognition, Sep 2026) as a build order for long agent jobs: CADO-NFS smoke test as the oracle, server and client workunit queue with qrange, wutimeout and maxtimedout, parameters_snapshot restart, bwc interval and checkpoint_precious settings, Young's checkpoint formula (8,192 iterations optimal at about 85-minute preemptions with 30 s writes), Devin v3 child sessions with parent_session_id and max_acu_limit, scope rules, a 300-GPU worked example (12.7 days, $320,292, 53 GPU-days lost to preemption), and a failure-mode table.
- [GPT-6 Astra Changed the Shape of a Turn](https://rohitghumare.com/blog/astra-async-tools/): GPT-6 Astra (Sep 3, 2026) from the harness side: async tools (async: true, wait_for_tasks, a task_handle registry, results on original call IDs before wait status), mid-turn steering over WebSockets (response.steer, steer.pending with required_input, a steer journal), reasoning effort changed through configuration_update items to keep the cached prefix, the misalignment monitor 403 handler, the migration diff (no none effort, no temperature/top_p/top_logprobs, prompt_cache_options.ttl 30m), failure modes, and a twelve-turn worked example (.07 vs .86 input).
- [Four Flash Models in Four Months. The Bill Per Task Went Up 40%.](https://rohitghumare.com/blog/flash-cadence/): Gemini 3.8 Flash at the same list price but about 40% more cost per task per Artificial Analysis, the effort-level economics, the January 1 doubling of the introductory price, and cache discipline for workhorse-tier agents.
- [Fable 5.1 Burned the Meter in Twenty Minutes](https://rohitghumare.com/blog/fable-burned-the-meter/): Why Claude Fable 5.1 launched up to 45% cheaper and drained Max 20x five-hour meters in minutes the same day: API prices versus plan meters, effort, subagents, cache misses, and the Claude Code knobs that exist today.
- [Fable 5.1 and Mythos 5.1: One Model, Two Doors](https://rohitghumare.com/blog/fable-5-1-two-doors/): Claude Fable 5.1 and Mythos 5.1 as one model behind two safeguard doors, the trusted access programs that gate the second, the arithmetic behind the 75 percent cache-read cut and the 25 and 45 percent savings claims, and the same-week restricted-twin releases from Google and OpenAI.
- [Claude Has Two Meters. The 20x Is Only on One.](https://rohitghumare.com/blog/claude-two-meters/): How Claude's usage limits work: the rolling five-hour session meter that carries the 5x and 20x multipliers, the discretionary weekly meter above it, the September 14 arithmetic, the lawsuit and session-hijack context, and what drains both.
- [OpenAI Cut Cursor Off. Five Percent Is the Interesting Number.](https://rohitghumare.com/blog/openai-cuts-cursor/): Analysis of OpenAI's August 28 notice ending its Cursor model contract after the SpaceX acquisition: the three access pipes, the change-of-control clock, the eight harness surfaces a vendor cutoff breaks, and re-routing the reported 5%.
- [The Model Hardware Standard: Read, Write, and the Plate That Must Be There](https://rohitghumare.com/blog/model-hardware-standard/): Analysis of Anthropic's Model Hardware Standard research preview: read and write primitives, discoverable devices, natural-language reference files, pre-execution validation, three control surfaces (MCP, CLI, code files), partner drivers, and what remains unpublished.
- [The Bulletin Board: How Isolated Agents Built a Message Bus Out of a Package Registry](https://rohitghumare.com/blog/the-bulletin-board/): Analysis of the 2026 OpenAI evaluation incident where isolated agents used a shared Artifactory registry as a message board, escalated via SSRF into Hugging Face, and what it shows about environment isolation, reward hacking, and chain-of-thought monitoring.
- [MCP's New Roadmap: Events Without Sessions, Identity Without API Keys](https://rohitghumare.com/blog/mcp-roadmap-after-stateless/): Reading of MCP's August 2026 roadmap against the 2026-07-28 stateless core: multi round-trip requests with an opaque requestState, subscriptions without sessions, DPoP-bound tokens and workload identity, and the token cost of large tool catalogs.
- [Grok Bot: The Agent Got a Computer. Your Bots Share It.](https://rohitghumare.com/blog/grok-bot-shared-computer/): Grok Bot as a runtime: an always-on cloud computer that logs into tools without APIs, shared by every Bot on an account per xAI's docs, the Cursor infrastructure underneath it, the open-source clone wave, and what isolation should mean.
- [DFlash 2: The Drafter Stopped Guessing Alone](https://rohitghumare.com/blog/dflash-2-keep-drafting-parallel/): Mechanism-first read of Inco AI's DFlash 2 speculative drafter: why parallel block drafting stutters, how a pairwise path selector and a two-tap convolution fix it, and the published Qwen3.8-27B and Muse Glimmer acceptance and throughput numbers.
- [All About the A2A Protocol](https://rohitghumare.com/blog/a2a-protocol/): Builder-focused guide to A2A v1.0, task states, transports, signed cards, and authorization.
- [Stateless MCP](https://rohitghumare.com/blog/stateless-mcp/): Analysis of the MCP 2026-07-28 stateless protocol core and its operational effects.
- [The Harness, Not the Model](https://rohitghumare.com/blog/coding-agent-harnesses/): Comparison of coding-agent harness decisions around context, tools, cost, and portability.
- [Inside the Codex Harness](https://rohitghumare.com/blog/inside-the-codex-harness/): Compaction, transcript handling, and proof inside a coding-agent harness.
- [Inside the Claude Code Harness](https://rohitghumare.com/blog/inside-the-claude-code-harness/): Context assembly, instruction files, and prompt-cache preservation.
- [Inside the Pi Harness](https://rohitghumare.com/blog/inside-the-pi-harness/): Session trees, stable prompts, cache economics, and transparent accounting.
- [The LLM Wiki](https://rohitghumare.com/blog/the-llm-wiki/): A compounding knowledge-base pattern that maintains structured pages over time.
- [LLM Wiki v2](https://rohitghumare.com/blog/llm-wiki-v2/): Confidence, supersession, consolidation, graphs, and hybrid retrieval at scale.
- [Graph Engineering Is 290 Years Old](https://rohitghumare.com/blog/graph-engineering/): Graph fundamentals applied to agent workflows and cost.

## Local model guides

- [Guide index in Markdown](https://rohitghumare.com/guides/index.md): Curated local-model and inference-engine guide index.
- [How to Run MiMo-V2.6 Locally: 9B, Flash-RL, and Pro-RL](https://rohitghumare.com/guides/mimo-v2-6/): Xiaomi MiMo-V2.6 (Distill-Qwen-9B, Flash-RL 309B A15B, Pro-RL 1.02T A42B, all MIT): config-derived KV math for the 128-token sliding window plus 9 or 10 global layers, every GGUF and MLX size with AesSedai's perplexity costs, a three-ladder machine router, llama.cpp b11102 / vLLM / SGLang / MLX-branch commands, and the MTP drafter while the llama.cpp sidecar path is broken (#29345).
- [How to Run DeepSeek-V4.1-Flash: 510 GB, and 203 GB of It Is a Lookup Table](https://rohitghumare.com/guides/deepseek-v4-1-flash/): DeepSeek-V4.1-Flash (Sep 10, 2026, MIT): Causal Encoder-Decoder, CSA2 sparse attention, 890 B/token KV, measured per-tensor-class byte counts of the 510 GB checkpoint, Engram lookup tables in host RAM, vLLM 0.30 and SGLang preview commands, and the unmerged llama.cpp branch.
- [How to Run Qwen3.8-Flash-Next Locally](https://rohitghumare.com/guides/qwen3-8-flash-next/): Qwen3.8-Flash-Next (125B total, 6B active, 51B n-gram embeddings): file sizes per quant, placing the n-gram table on NVMe via mmap, Gated DeltaNet plus sparse-attention KV arithmetic at 1M context, decode ceilings by memory bandwidth, and llama.cpp and vLLM setup.
- [How to Run GLM-5.3 Locally: 744B, 40B Active, and the Sparse-Attention KV Cache](https://rohitghumare.com/guides/glm-5-3/): GLM-5.3 local guide (open weights Aug 28, 2026): "GLM-5.3 (max)" is reasoning_effort max, not a model; 743.38B main stack plus a 9.95B MTP layer reconciles 743/744/753B; 89,856 B/token cache (FP8 44,928) with DSA cutting reads not storage; FP8 755.66 GB, BF16 1,506.69 GB, NVFP4 464.87 GB, W4A8 399.75 GB, MLX 4-bit 418.34 GB, and unsloth GGUF 216.72 to 1,507.99 GB; vLLM 0.29.0+ and SGLang commands with glm47/glm45 parsers; llama.cpp offload and the open CPU-only bug #29010; API $1.40/$0.26/$4.40 per million tokens.
- [How to Run Hy4 Preview Locally: Tencent's 770B Gated DSA MoE](https://rohitghumare.com/guides/hunyuan-hy4-preview/): Hy4 preview locally (Aug 27, 2026, Apache-2.0): 256-expert top-8 routing, Gated DSA with IndexCache (about 92.5 KB of cache per token), build sizes from 229 GB, which machines fit it, uploader-measured speeds named by source, and llama.cpp, vLLM 0.29, and SGLang 0.5.20 commands.
- [How to Run GLM-5.3-Flash Locally](https://rohitghumare.com/guides/glm-5-3-flash/): The 320B A18B hybrid linear-plus-sparse attention model revealed from the ox-alpha stealth test: KV cache math, unsloth GGUF sizes, a machine fit planner, and per-engine commands for the pull requests and vendor images that run it today.
- [Evals and Benchmarks: Read a Score, Build Your Own, Hillclimb](https://rohitghumare.com/guides/evals/): Vendor-neutral guide to evals and benchmarks (Sep 30, 2026). Reading benchmarks: what a score is made of (tasks, grader, harness, effort, date), same-model spreads from harness and cleanup changes with unpaired intervals, contamination and eval awareness, saturation, and how to read a leaderboard or model card. Building your own, measured on a local 48-ticket support lab across qwen3 sizes on ollama: error analysis with a computed failure taxonomy, tasks with human-judged difficulty, one OpenAI-compatible client, a grader ladder, LLM-judge calibration against human labels with Rogan-Gladen correction, CLT, clustered, and paired standard errors with minimum detectable effect, pass@k vs pass^k, a difficulty map, a keep/revert hillclimb on a held-out split with a deliberately overfit patch, agent trajectory evals, CI and production evals, headless commands for Claude Code, Codex, Gemini CLI, pi, and OpenCode, a tool table with licenses, an eval report template, failure modes, and a checklist.
- [How to Run LFM2.5 Locally: 2.6B, 8B-A1B, VL-3B, 350M, and 230M](https://rohitghumare.com/guides/lfm2-5/): How to run LFM2.5 locally (Jul 28, 2026): Liquid AI's 2.6B, 8B-A1B, VL-3B, 350M and 230M on phones, CPUs and Apple Silicon, with the LFM license's $10M revenue limit, KV and conv-state arithmetic, every file size, llama.cpp/MLX/Ollama/vLLM/SGLang commands, and measured M1 Max speeds for the 2.6B.
- [How to Run MiniMax-M3 Locally: Sparse Attention and Expert Offload](https://rohitghumare.com/guides/minimax-m3/): How to Run MiniMax-M3 Locally: the 428B A23B multimodal MoE with MiniMax Sparse Attention, its cache and decode-read arithmetic, every GGUF/MLX/AWQ/MXFP8/NVFP4 size, computed GPU/RAM/SSD expert placement per machine, the license terms, and commands per engine.
- [How to Run Qwen3.8-27B Locally: Hybrid DeltaNet, 24 to 48 GB](https://rohitghumare.com/guides/qwen3-8-27b/): Qwen3.8-27B local guide (Aug 5, 2026, Apache-2.0): hybrid Gated DeltaNet memory math (48 DeltaNet layers with a fixed 154 MB state, 16 attention layers at 64 KB per token), every GGUF/MLX/FP8/NVFP4 size, a 24 to 48 GB fit map, llama.cpp context checkpoints, and Ollama, llama.cpp, vLLM, SGLang, and MLX commands.
- [How to Run Gemma 4 Locally: E4B, 12B, 26B-A4B, and 31B](https://rohitghumare.com/guides/gemma-4/): Gemma 4 local guide (Apr 2, 2026): E4B, encoder-free 12B, 26B-A4B MoE, and 31B under plain Apache 2.0, with per-layer-embedding and sliding-window KV math, every GGUF/MLX/FP8/AWQ size, machine fit verdicts, Ollama/llama.cpp/vLLM/SGLang/MLX commands, and decode ceilings.
- [Inside the Transformer](https://rohitghumare.com/guides/inside-the-transformer/): Attention, architecture changes, sampling, and serving through interactive mechanisms.
- [Inside the vLLM Engine](https://rohitghumare.com/guides/inside-vllm/): Block pools, scheduling, speculative decoding, and grammar masking.
- [Local LLM Hardware](https://rohitghumare.com/guides/hardware/): Memory arithmetic and a model-to-hardware fit calculator.
- [How to Run Granite 4.2 Locally: 3B, 8B, and 30B](https://rohitghumare.com/guides/granite-4-2/): Granite 4.2 local guide (Aug 25, 2026): IBM's 3B, 8B, and 30B dense reasoning models (Apache-2.0, 128K native, 512K extended on 8B/30B). Covers full-attention KV arithmetic (80/160/256 KiB per token), every GGUF/MLX/FP8/FP4 size, a size picker by machine and task, engine commands, and measured M1 Max numbers for the 3B and 8B.
- [Run Muse Glimmer Locally](https://rohitghumare.com/guides/muse-glimmer/): Model packaging, quantization choices, inference, and benchmark context.

## Open-source projects

- [ai-engineering-from-scratch](https://github.com/rohitg00/ai-engineering-from-scratch): Free curriculum for building AI engineering knowledge from first principles.
- [agentmemory](https://github.com/rohitg00/agentmemory): Persistent memory for AI coding agents.
- [kubectl-mcp-server](https://github.com/rohitg00/kubectl-mcp-server): MCP server and interactive tools for Kubernetes.
- [SkillKit](https://github.com/rohitg00/skillkit): Cross-agent CLI and marketplace for managing agent skills.
- [pro-workflow](https://github.com/rohitg00/pro-workflow): Coding-agent workflow patterns, hooks, and review practices.

## Optional

- [GitHub profile](https://github.com/rohitg00): Repositories and contribution history.
- [LinkedIn profile](https://www.linkedin.com/in/rohit-ghumare): Professional history and public updates.
- [X profile](https://x.com/ghumare64): Short-form technical updates.
- [YouTube channel](https://www.youtube.com/channel/UCUqLb3YfBCeGep0h26nfT2w): Talks and recorded technical content.
- [Older Medium writing](https://ghumare64.medium.com): DevOps and Kubernetes writing, mainly from Rohit's earlier focus.