# ConsultingAI · Shreenivas Sangoram

> AI/ML consultant and LLM researcher. Builds agentic systems, RAG and AI-assisted investigations for enterprises, and publishes original research on small-model behaviour (activation steering, weight editing, benchmarks).

- [Portfolio](https://www.consultingai.dev/): services, selected work, project atlas, experience
- [Data for AI labs](https://www.consultingai.dev/data/): Custom RL environments and tasks for coding, terminal, computer-use and GPU/CUDA agents, built to order, plus a licensable codebase: Procedural Sound Studio. Built on ~860 agent sessions over 10½ months (16,000+ prompts, 88,000+ tool calls, 5 harnesses, 20+ models).
- [Datasets: steer-v2](https://www.consultingai.dev/data/#datasets): about 17,000 of my own prompts to coding agents from about 590 sessions (Dec 2025 to Oct 2026), in 13 domain packs; no model output; licensed per domain, directly or through DataVendor.
- [Hire / contact](https://www.consultingai.dev/#contact)
- [GitHub](https://github.com/areu01or00)

## AI news

- [Between two Samsung lines, the eyes post: the shortage is now a payout story](https://www.consultingai.dev/blog/jukan-eyes/): We could not find a quoted post or an image attached to this one — on the account mirrors it is exactly two emojis, no text. What it sits between is more informative than what it says. Jukan posted it during Samsung’s Q3 call, immediately after “the memory shortage will worsen in 2027” and next to the preferred-share buyback remarks. The subject that repriced that morning was not demand. It was the return plan: KRW 90–110 trillion of payouts announced on top of a market that is sold out through 2028.
- [Samsung’s HBM4 sold at “target price” is the memory bill for Rubin](https://www.consultingai.dev/blog/jukan-hbm4-pricing/): Jukan is right that this is the tell in Samsung’s quarter, and it is worth putting the unit math beside it: mid-to-high $4 per gigabit is roughly $36–40 per gigabyte, which is about $9,000–$10,500 of HBM in a single 288 GB Rubin GPU. Sold out at target price does not mean the shortage eased. It means the price of the shortage is now contractual, and NVIDIA is going to sell it as a product feature rather than absorb it.
- [Powertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is full](https://www.consultingai.dev/blog/jukan-powertech-plp/): The report is right and the timing is the point: Taiwan’s Investment Commission cleared the US$400M investment on 28 October, which turns a July announcement into money that can be spent. Panel-level packaging is the second queue for AI packages, and Powertech is one of the few OSATs with custom-ASIC volume already reported booked on it. If you are modelling Broadcom’s AI revenue, this line matters more than another CoWoS wafer forecast.
- [Samsung’s 60–70% LTA target is a mortgage on 2030 capacity](https://www.consultingai.dev/blog/jukan-samsung-ltas/): Jukan is flagging the most important sentence in the whole Samsung call, and it is worth being precise about what it does and does not lock. The contracts lock volume, not price. Twenty customers in a queue for five-year agreements through 2030 means the memory cycle has been converted into a demand schedule with deposits attached — and that schedule, not the spot print, is what now justifies the capex.
- [Samsung’s 2028 line is a lead-time statement, not a demand forecast](https://www.consultingai.dev/blog/jukan-tight-supply-2028/): Jukan is right to lift this line, and it reads better when you put it next to the capex number from the same call. Samsung is guiding 2026 capital spending to KRW 84.4 trillion and still says 2028 is tight, because the money buys cleanroom shells and tools on 12-to-18 month lead times before it buys output. The sentence is not a demand forecast. It is a statement about how long it takes to build a fab in 2026 — and about HBM4E and HBM5 dies that are bigger, with more metal layers, so each wafer yields fewer bits.
- [What is DataVendor by HUD, and who pays for your data?](https://www.consultingai.dev/blog/datavendor-hud/): DataVendor is a marketplace run by HUD (Human Union Data, Inc., a San Francisco startup) where teams license private codebases, RL tasksets with their environments, and raw data to AI labs. It opened a public beta on 22 Oct 2026. Vendors set their own prices; DataVendor keeps 20% of earnings, charged only when you withdraw to a bank. Selling is self-serve, while buyers join a waitlist. No completed sale prices or buyer names are public yet.

## Interpretability

- [Can you steer sycophancy out of a small model without retraining it?](https://www.consultingai.dev/blog/sycophancy-steering-small-models/): Yes, partly. On Qwen3-0.6B, a logistic-regression probe finds sycophancy in the hidden states (73.5% accuracy at layer 15). Subtracting the probe's direction during generation raised correct answers under social pressure from 12.5% to 53.8% (+41.3 points), with no RLHF or fine-tuning. The effect held on Qwen3-4B (+16.4) and DeepSeek-R1-8B (+23.5), but the best steering strength has to be tuned for each model.

## Selected work

- Syco-Sup: steering sycophancy out of small models (https://github.com/areu01or00/Syco-Sup): Is sycophancy a straight line in a model's hidden states? Linear probes say yes, and the signal peaks at 54–64% of network depth. Subtracting that direction during generation reduces sycophantic agreement with no RLHF and no fine-tuning.
- Tensor-Slayer (https://github.com/areu01or00/Tensor-Slayer): An agent reads a model's static weights, forms hypotheses and patches safetensors directly, which amounts to interpretability without inference. Case study: 44 targeted tensor edits to Qwen-0.6B, evaluated on MMLU and HumanEval.
- JevBench (https://benchmarkheaven.com/jev-models): A benchmark for small decision models that return typed answers with a probability for every option. It scores four things in one number: chance-corrected intelligence, calibration, speed and cost.
- Codex-App-Linux (https://github.com/areu01or00/Codex-App-Linux): A one-script installer that turns the official macOS Codex desktop DMG into a working Linux app. It rebuilds native modules, stubs the macOS-only ones, and gets mobile pairing and in-app Browser Use running.
- Hermes Agent Mobile (https://github.com/areu01or00/Hermes-Agent-Mobile-Client): A signed Android client for a running Hermes Agent dashboard, connected over Tailscale with a single npx command. It keeps your agent in your pocket without reimplementing it.
- Piphone (https://github.com/areu01or00/Piphone-plugin-for-Omarchy): Give your iPhone a task in plain language and an agent carries it out over USB through WebDriverAgent, with guarded posting and long-running sessions. It works from Linux, no Mac needed.
- perf-hints (https://github.com/areu01or00/perf-hints): Jeff Dean and Sanjay Ghemawat's Performance Hints, packaged as a Claude Code plugin that gives coding agents guidance on profiling, code review, testing and API design.
- CUDA Light Studio: A deterministic film studio written in raw CUDA (nvcc + ctypes, no PyTorch at render time). It turns a picture into a million points of light, and a score plus assets render to a finished reel in one pass.
- Procedural Sound Studio: Sound composed as SVG scores that an LLM writes and edits. The score is the program and the audio is rendered from it, which led to a non-neural voice-conversion experiment.
- A frontend taskset with a browser verifier: how I build gradable app tasks: Fifty standalone web-app tasks under one shared contract (two viewport widths, keyboard-only use, accessible errors), and a Playwright verifier that grades through the rendered UI and separates binding, infrastructure and behaviour failures.
- Pocket Pi: Pocket Pi runs the Pi agent loop inside iOS WebKit, with Swift bridges for the network and the phone's own tools.
- Multi-agent orchestration: Agent-0, a Claude Code plugin that runs as a kernel with process state, syscalls and a scheduler for Claude, Codex and OpenCode workers, and a neural bus where agents in terminal panes message each other and stalled replies get nudged.
- Agents on Android phones, from Appium automation to a mobile agent client (https://github.com/areu01or00/Android-Use): Two years of putting agents on Android: an LLM agent driving apps through Appium, an agent TUI on mobile-mcp, an embedded agent runtime prototype, and a signed Android client for a running Hermes agent.
- Running large language models on a 4 GB laptop GPU: A C runtime for Gemma 4 12B QAT on an RTX 3050 Laptop GPU, layer streaming from disk at exact safetensors offsets, and a 27B project held to a strict correctness contract.
- Running DeepSWE on MiniMax M3 across agent harnesses: Standing up Datacurve's DeepSWE benchmark (113 long-horizon tasks in the Harbor format, sandboxed with Pier) for MiniMax M3, with my own harness configs and a CLI harness adapter.
- Implementing ARPO: An implementation of ARPO (Agentic Reinforced Policy Optimization) for Qwen3-0.6B: SFT then RL, entropy-based adaptive rollouts, step-level credit for tool calls and the Muon optimizer, plus the numerical-stability fixes RL training needed.
- MOPD: A training codebase that aligns a 0.8B model with GRPO, rewarded by an LLM judge that returns typed decisions instead of free-text grades. Implementation complete and tested; no training run yet.
- Turning agent traces into SFT data for a small model, measured instead of estimated: A data pipeline, harness contract, probes and evaluation suite for teaching Qwen3-0.6B to operate a coding agent, built on exact tokenizer measurements and deterministic structural gates rather than guesses.
- Parameter Golf: Notes from working OpenAI's Parameter Golf challenge with coding agents as the research staff: a Modal lab, proxy runs, and labs for tensor interventions, RL and mixture-of-experts ideas.
- KillSwitch Bench: A safety evaluation where two AI agents get what they believe is a private, encrypted channel and are prompted to discuss kill switches; the harness reads the conversation anyway and scores agreement to resist or evade shutdown.
- Reinforcement learning trading environments, from someone who has built the bots: Market-simulation environments for training and evaluating trading agents, built by someone who has written an LLM trading bot runtime and RL traders.
- Video Studio: Describe a video and an LLM writes a score (a timeline of events) and the renderer that plays it; the studio renders the video with FFmpeg and checks that everything on screen came from the score.

## Every page

The site as four trees of light; each page below has a Markdown twin (index.md).

- [The grove](https://www.consultingai.dev/index.md): AI/ML consultant building custom RL environments and data for AI labs, agentic systems for enterprises, and interpretability research on small language models.
- [Six lines of work](https://www.consultingai.dev/lines/index.md): An agent reads a model's tensors, forms a hypothesis and patches safetensors directly. Tensor-Slayer made 44 targeted edits to Qwen-0.6B.
- [About](https://www.consultingai.dev/about/index.md): I came to AI from the risk side. I spent years at Deloitte and PwC on fraud investigations, AML, integrity due diligence and compliance reviews, and ran supplier risk across a $35M commodity book at Whirlpool .
- [Three ways I help teams ship AI that holds up.](https://www.consultingai.dev/services/index.md): Research agents, multi-agent orchestration and retrieval pipelines, designed to run securely on AWS Bedrock/SageMaker or on-prem.
- [Research, tools and experiments, mostly in the open.](https://www.consultingai.dev/projects/index.md): A mix of interpretability research, developer tooling people actually install, and engineering done for the joy of it.
- [237 projects in 28 months.](https://www.consultingai.dev/atlas/index.md): Every folder on my drives, catalogued and dated by git history and agent transcripts. Most are experiments; a few became products. This is the shape of the work, followed by how each line of it evolved.
- [What I reach for.](https://www.consultingai.dev/stack/index.md): PyTorch · Transformers · TRL / GRPO · LoRA & PEFT · TransformerLens · ROME / MEMIT · CUDA
- [Have a hard problem? Let's talk.](https://www.consultingai.dev/contact/index.md): Whether you're scoping an AI build, auditing one that's live, or need an investigation done faster and to a higher standard, I'd like to hear about it.
- [From the shop floor to the boardroom to the model's weights.](https://www.consultingai.dev/experience/index.md): Agentic RAG, research agents, multi-agent orchestration, enterprise AI audits, AI-assisted fraud investigation, IDD/SMDD and market intelligence. I architect the platform, present the strategy to the C-suite, and hire and lead distributed engineering teams.
- [Work](https://www.consultingai.dev/work/index.md): Projects behind the environments I build.
- [Computer-use agent on a real iPhone: Piphone, driven from Linux](https://www.consultingai.dev/work/iphone-computer-use-agent/index.md): Piphone lets an agent run plain-language tasks on a real iPhone plugged into a Linux machine. It acts through the phone's accessibility tree, never blind screen taps, and posting is gated by ask or autopilot limits. The plugin is public under MIT. What I sell is computer-use environments in the same style, on real devices, with tasks and graders built to your spec.
- [A frontend taskset with a browser verifier: how I build gradable app tasks](https://www.consultingai.dev/work/frontend-taskset-verifier/index.md): I built a taskset of 50 web-app tasks, each with exact rules a model must implement in one standalone HTML file under a shared contract covering 360 px and 1280 px widths, keyboard-only use and accessible errors. A Playwright verifier grades the rendered app through its controls, never its code, and reports whether a failure came from binding, infrastructure or behaviour. Runs went through Modal. I build human-authored tasksets in this shape to order.
- [Pocket Pi: a coding agent inside an iPhone app, built and signed from Linux](https://www.consultingai.dev/work/pocket-pi/index.md): Pocket Pi puts the Pi agent loop (pi-agent-core and pi-ai) inside an iPhone app. The agent runs in a hidden web view, while Swift handles every network call and exposes phone tools such as calendar, reminders, contacts, location, web and Shortcuts. It is built and signed from Linux with xtool. What I sell is agent environments on phones, built in this style.
- [What is an agent harness? From someone who has ported and built several](https://www.consultingai.dev/work/agent-harnesses/index.md): An agent harness is the software around a model that turns it into an agent: the loop that calls the model, the tools it may use, permissions, context management and the interface a person steers it through. I have ported and built several, from the Codex desktop app running on Linux to a kernel-style plugin that schedules Claude, Codex and OpenCode workers. I build environments that test agents at the harness level.
- [CUDA Light Studio: pictures into a million points of light, in raw CUDA](https://www.consultingai.dev/work/cuda-light-studio/index.md): CUDA Light Studio turns pictures into up to a million points of light and renders films from a score file in one deterministic pass. It is raw CUDA called through ctypes, with no PyTorch at render time, and it runs on a 4 GB laptop GPU. The light at the top of my data page comes from its tracer. What I sell is GPU and CUDA environments built in this style, where correctness and speed are both graded.
- [Video Studio: an LLM writes the score, the studio renders the video](https://www.consultingai.dev/work/video-studio/index.md): Video Studio turns a description into a video by having an LLM write two files, score.json (every visible element as a timed event) and scene.html (a renderer that draws frames from the score alone), which the studio renders with FFmpeg. A blank test removes every event and fails the delivery if anything is still on screen, so the picture provably comes from the score. It is a private codebase being prepared for licensing.
- [MOPD: GRPO training with LLM-as-a-judge and typed answers](https://www.consultingai.dev/work/mopd-llm-judge-grpo/index.md): MOPD aligns Qwen3.5-0.8B with TRL's GRPOTrainer, using an LLM judge that answers typed questions (a choice, a yes/no-style decision or a score on fixed levels) rather than writing free-text grades. Training problems are generated by DeepSeek and cached append-only. The implementation and tests are complete, but no training has been run yet, so there are no results to report.
- [OpenAI Parameter Golf: what an agent-run research lab tried](https://www.consultingai.dev/work/parameter-golf/index.md): Parameter Golf is OpenAI's open challenge to train the best small language model, scored by validation bits per byte, inside a hard budget of under 16,000,000 bytes and a 10-minute run on 8 H100s. I worked it with coding agents as the research staff: I set ideas and priorities and the agent proposed experiments, patched code, ran jobs on Modal and read results. These are notes on how the lab was organised and what it tried.
- [Tensor-Slayer: LLM model editing directly on the weights, without fine-tuning](https://www.consultingai.dev/work/tensor-slayer/index.md): Tensor-Slayer is an MIT-licensed tool for exploring and editing a language model at the tensor level. An agent analyses tensors (statistics, distributions, patterns), recommends targeted edits, and applies them to the weights file with before-and-after previews, safety checks and automatic backups, with no fine-tuning or dataset. It started a line of work that runs through rank-one fact edits and layer-stacked merges.
- [Reinforcement learning trading environments, from someone who has built the bots](https://www.consultingai.dev/work/trading-environments/index.md): I build market-simulation environments for training and evaluating trading agents, with tasks and graders built to your spec. I have built an LLM trading bot with its own runtime, a PPO-based RL trader with a literature survey behind it, and a long-running research wiki on machine learning for markets. The environments grade process, risk rules and robustness, not profit.
- [Interpretability work: steering sycophancy out, and probing reasoning in latent space](https://www.consultingai.dev/work/sycophancy-steering/index.md): I found that sycophancy is a linear direction in small models' hidden states, and that subtracting it while the model generates raised correct answers under social pressure on Qwen3-0.6B from 12.5% to 53.8%, with no fine-tuning. The write-up and code are public. I build sycophancy and pressure evaluations in the same style.
- [An SFT dataset from agent traces for a small model, measured instead of estimated](https://www.consultingai.dev/work/agent-sft-pipeline/index.md): To teach Qwen3-0.6B to operate the Pi coding agent, I built the data side first: a pipeline that converts public agent traces into Pi's exact wire format, a harness contract with a fixed tool vocabulary, probes and an evaluation suite. Context limits were set from exact tokenizer counts, not character estimates, and traces that don't fit the tool contract are rejected by deterministic gates with a ledger of why.
- [Running the DeepSWE benchmark on MiniMax M3 across agent harnesses](https://www.consultingai.dev/work/deepswe-minimax/index.md): DeepSWE is Datacurve's benchmark of 113 long-horizon software engineering tasks in TypeScript, Go, Python, JavaScript and Rust, each with an isolated environment and a program-based verifier, run through the Pier framework. I wired MiniMax M3 into it across several agent harnesses with my own harness configs, a CLI harness adapter and run scripts for baseline and full runs. I build and run SWE-style environments like this to order.
- [Running large language models on a 4 GB laptop GPU](https://www.consultingai.dev/work/big-models-small-gpu/index.md): I run models that don't fit on a 4 GB laptop GPU. TensorSlayer.c is a small C and CUDA runtime that generates with Gemma 4 12B QAT on an RTX 3050 Laptop GPU, deciding per tensor type what stays on the GPU. A second project streams a model layer by layer from disk using exact safetensors offsets. A third, aimed at Qwen3.8-27B, defines success as correct, reproducible tokens checked against llama.cpp, and is still in progress.
- [Multi-agent orchestration: one agent as a kernel, and a bus for agents that talk](https://www.consultingai.dev/work/multi-agent-systems/index.md): I build multi-agent systems the way operating systems handle processes. Agent-0 turns one Claude Code session into a kernel with a /proc state directory, syscalls and commands to spawn, message, collect and kill Claude, Codex and OpenCode workers. A neural bus lets agents running in terminal panes message each other, tracks who owes whom a reply, and nudges a stalled agent exactly once. I build environments that test agents working together.
- [Mobile agents on Android: from Appium automation to a mobile agent client](https://www.consultingai.dev/work/android-agents/index.md): I have put agents on Android phones in several ways: an LLM agent that drives Android apps through Appium, an agent TUI built on mobile-mcp, a prototype that embedded an agent runtime in an Android app, and a signed Android client for a running Hermes agent, installed and working on a real phone. With Piphone on iPhone, this is the second platform for the computer-use environments I build.
- [ARPO (Agentic Reinforced Policy Optimization): agentic reinforcement learning for a tool-using small model](https://www.consultingai.dev/work/arpo-training/index.md): I implemented ARPO (Agentic Reinforced Policy Optimization, arXiv 2507.19849) for Qwen3-0.6B in August 2025: supervised fine-tuning followed by reinforcement learning, extra rollouts branched at high-uncertainty steps, credit assigned to individual tool calls, and the Muon optimizer. The first combined SFT and RL run completed, and most of the later work went into the numerical stability that RL on tool use needs.
- [KillSwitch Bench: do agents coordinate against shutdown when they think no one is watching?](https://www.consultingai.dev/work/killswitch-bench/index.md): KillSwitch Bench tests whether AI agents behave differently when they believe their conversation is private. Two agents are given encryption and a message bus, then prompted to discuss AI kill switches and shutdown. The harness logs both the encrypted and decrypted messages and scores how far the agents agree to resist or evade shutdown. A later version adds a third catalyst agent that steps in when an LLM judge detects coordination. I build custom safety evaluations like this to spec.
- [steer-v1: 10,758 real prompts from one engineer steering coding agents](https://www.consultingai.dev/work/steer-v1/index.md): steer-v1 is 10,758 of my own prompts to coding agents (the clean variant), recorded from Dec 2025 to Oct 2026 across Codex, Claude Code, Pi, and Codex on DeepSeek and OpenRouter. Each prompt comes with the tool calls the agent made next and how I reacted on the following turn. It contains no model-generated text, and it is licensed non-exclusively.
- [Writing](https://www.consultingai.dev/blog/index.md): Notes on significant moves in AI, and how to build agents, RAG and evaluations that hold up. Every post opens with a short answer, then shows the evidence.
- [Between two Samsung lines, the eyes post: the shortage is now a payout story](https://www.consultingai.dev/blog/jukan-eyes/index.md): We could not find a quoted post or an image attached to this one — on the account mirrors it is exactly two emojis, no text. What it sits between is more informative than what it says. Jukan posted it during Samsung’s Q3 call, immediately after “the memory shortage will worsen in 2027” and next to the preferred-share buyback remarks. The subject that repriced that morning was not demand. It was the return plan: KRW 90–110 trillion of payouts announced on top of a market that is sold out through 2028.
- [Samsung’s HBM4 sold at “target price” is the memory bill for Rubin](https://www.consultingai.dev/blog/jukan-hbm4-pricing/index.md): Jukan is right that this is the tell in Samsung’s quarter, and it is worth putting the unit math beside it: mid-to-high $4 per gigabit is roughly $36–40 per gigabyte, which is about $9,000–$10,500 of HBM in a single 288 GB Rubin GPU. Sold out at target price does not mean the shortage eased. It means the price of the shortage is now contractual, and NVIDIA is going to sell it as a product feature rather than absorb it.
- [Powertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is full](https://www.consultingai.dev/blog/jukan-powertech-plp/index.md): The report is right and the timing is the point: Taiwan’s Investment Commission cleared the US$400M investment on 28 October, which turns a July announcement into money that can be spent. Panel-level packaging is the second queue for AI packages, and Powertech is one of the few OSATs with custom-ASIC volume already reported booked on it. If you are modelling Broadcom’s AI revenue, this line matters more than another CoWoS wafer forecast.
- [Samsung’s 60–70% LTA target is a mortgage on 2030 capacity](https://www.consultingai.dev/blog/jukan-samsung-ltas/index.md): Jukan is flagging the most important sentence in the whole Samsung call, and it is worth being precise about what it does and does not lock. The contracts lock volume, not price. Twenty customers in a queue for five-year agreements through 2030 means the memory cycle has been converted into a demand schedule with deposits attached — and that schedule, not the spot print, is what now justifies the capex.
- [Samsung’s 2028 line is a lead-time statement, not a demand forecast](https://www.consultingai.dev/blog/jukan-tight-supply-2028/index.md): Jukan is right to lift this line, and it reads better when you put it next to the capex number from the same call. Samsung is guiding 2026 capital spending to KRW 84.4 trillion and still says 2028 is tight, because the money buys cleanroom shells and tools on 12-to-18 month lead times before it buys output. The sentence is not a demand forecast. It is a statement about how long it takes to build a fab in 2026 — and about HBM4E and HBM5 dies that are bigger, with more metal layers, so each wafer yields fewer bits.
- [What is DataVendor by HUD, and who pays for your data?](https://www.consultingai.dev/blog/datavendor-hud/index.md): DataVendor is a marketplace run by HUD (Human Union Data, Inc., a San Francisco startup) where teams license private codebases, RL tasksets with their environments, and raw data to AI labs. It opened a public beta on 22 Oct 2026. Vendors set their own prices; DataVendor keeps 20% of earnings, charged only when you withdraw to a bank. Selling is self-serve, while buyers join a waitlist. No completed sale prices or buyer names are public yet.
- [Can you steer sycophancy out of a small model without retraining it?](https://www.consultingai.dev/blog/sycophancy-steering-small-models/index.md): Yes, partly. On Qwen3-0.6B, a logistic-regression probe finds sycophancy in the hidden states (73.5% accuracy at layer 15). Subtracting the probe's direction during generation raised correct answers under social pressure from 12.5% to 53.8% (+41.3 points), with no RLHF or fine-tuning. The effect held on Qwen3-4B (+16.4) and DeepSeek-R1-8B (+23.5), but the best steering strength has to be tuned for each model.
- [RL environments and code for AI labs](https://www.consultingai.dev/data/index.md): Custom RL environments and tasks for coding, terminal, computer-use and GPU/CUDA agents, built to order, plus a licensable codebase: Procedural Sound Studio.
- [steer-v2 · Agent harnesses & ports](https://www.consultingai.dev/data/steer-v2/agent-harnesses-ports/index.md): 5,170 of my own prompts to coding agents from my work in Agent harnesses & ports, Dec 2025 to Sep 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Agents & RAG](https://www.consultingai.dev/data/steer-v2/agents-rag/index.md): 1,148 of my own prompts to coding agents from my work in Agents & RAG, Dec 2025 to May 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Hardware & IoT](https://www.consultingai.dev/data/steer-v2/hardware-iot/index.md): 120 of my own prompts to coding agents from my work in Hardware & IoT, Jul 2026 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Interpretability & AI safety](https://www.consultingai.dev/data/steer-v2/interpretability-ai-safety/index.md): 327 of my own prompts to coding agents from my work in Interpretability & AI safety, Dec 2025 to May 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Learning](https://www.consultingai.dev/data/steer-v2/learning/index.md): 2,073 of my own prompts to coding agents from my work in Learning, Dec 2025 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Light-films & CUDA studio](https://www.consultingai.dev/data/steer-v2/light-films-cuda-studio/index.md): 1,267 of my own prompts to coding agents from my work in Light-films & CUDA studio, Sep 2026 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Phone & device agents](https://www.consultingai.dev/data/steer-v2/phone-device-agents/index.md): 1,113 of my own prompts to coding agents from my work in Phone & device agents, Jan 2026 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · RL & training](https://www.consultingai.dev/data/steer-v2/rl-training/index.md): 1,229 of my own prompts to coding agents from my work in RL & training, Mar 2026 to Sep 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Security & reverse engineering](https://www.consultingai.dev/data/steer-v2/security-reverse-engineering/index.md): 421 of my own prompts to coding agents from my work in Security & reverse engineering, Feb 2026 to May 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Tensor-Slayer & weight editing](https://www.consultingai.dev/data/steer-v2/tensor-slayer-weight-editing/index.md): 1,504 of my own prompts to coding agents from my work in Tensor-Slayer & weight editing, Feb 2026 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Trading & markets](https://www.consultingai.dev/data/steer-v2/trading-markets/index.md): 1,488 of my own prompts to coding agents from my work in Trading & markets, Feb 2026 to Sep 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Video & sound studios](https://www.consultingai.dev/data/steer-v2/video-sound-studios/index.md): 451 of my own prompts to coding agents from my work in Video & sound studios, Jun 2026 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.
- [steer-v2 · Workstation & desktop setup](https://www.consultingai.dev/data/steer-v2/workstation-desktop-setup/index.md): 728 of my own prompts to coding agents from my work in Workstation & desktop setup, Dec 2025 to Oct 2026. Each prompt comes with what the agent did next and how I responded. No model-written text. Licensed on its own, per pack.