---
title: The grove
url: https://www.consultingai.dev/
original: https://www.consultingai.dev/
---


# Enterprise AI, built with an investigator's eye.

AI/ML Consultant · LLM Researcher

I'm Shreenivas Sangoram. I design agentic systems, RAG pipelines and AI-assisted investigations for firms where a wrong answer is expensive, and I open up small language models to see why they behave the way they do.

Start a conversation →

See the work

Data for AI labs →

9+

years in advisory & AI

237

projects built since mid-2024

72

client engagements: diligence & investigations

+41pp

sycophancy fix, no retraining

## Home

- [Six lines of work](/lines/)
- [About](/about/)
- [What I do](/services/)
- [Selected work](/projects/)
- [Atlas](/atlas/)
- [Toolkit](/stack/)
- [Contact](/contact/)
- [Experience](/experience/)

Home is a tree of light and its roots, drawn with Grok Imagine for this site.

## Work

- [Computer-use agent on a real iPhone](/work/iphone-computer-use-agent/)
- [A frontend taskset with a browser verifier](/work/frontend-taskset-verifier/)
- [Pocket Pi](/work/pocket-pi/)
- [What is an agent harness](/work/agent-harnesses/)
- [CUDA Light Studio](/work/cuda-light-studio/)
- [Video Studio](/work/video-studio/)
- [MOPD](/work/mopd-llm-judge-grpo/)
- [OpenAI Parameter Golf](/work/parameter-golf/)
- [Tensor-Slayer](/work/tensor-slayer/)
- [Reinforcement learning trading environments, from someone who has built the bots](/work/trading-environments/)
- [Interpretability work](/work/sycophancy-steering/)
- [An SFT dataset from agent traces for a small model, measured instead of estimated](/work/agent-sft-pipeline/)
- [Running the DeepSWE benchmark on MiniMax M3 across agent harnesses](/work/deepswe-minimax/)
- [Running large language models on a 4 GB laptop GPU](/work/big-models-small-gpu/)
- [Multi-agent orchestration](/work/multi-agent-systems/)
- [Mobile agents on Android](/work/android-agents/)
- [ARPO (Agentic Reinforced Policy Optimization)](/work/arpo-training/)
- [KillSwitch Bench](/work/killswitch-bench/)
- [steer-v1](/work/steer-v1/)

Work is a Lichtenberg figure: lightning frozen in acrylic · Timothee Rolland · CC BY-SA 4.0. [source](https://commons.wikimedia.org/wiki/File:Lichtenberg_Fractal_Figure_03.jpg)

## Writing

- [Between two Samsung lines, the eyes post](/blog/jukan-eyes/)
- [Samsung’s HBM4 sold at “target price” is the memory bill for Rubin](/blog/jukan-hbm4-pricing/)
- [Powertech’s $400M panel fab is where Broadcom’s ASIC volume goes when CoWoS is full](/blog/jukan-powertech-plp/)
- [Samsung’s 60–70% LTA target is a mortgage on 2030 capacity](/blog/jukan-samsung-ltas/)
- [Samsung’s 2028 line is a lead-time statement, not a demand forecast](/blog/jukan-tight-supply-2028/)
- [What is DataVendor by HUD, and who pays for your data](/blog/datavendor-hud/)
- [Can you steer sycophancy out of a small model without retraining it](/blog/sycophancy-steering-small-models/)

Writing is seed pods drifting down a thread of light · Stockcake · Stockcake (public domain).

## Data

- [Agent harnesses & ports](/data/steer-v2/agent-harnesses-ports/)
- [Agents & RAG](/data/steer-v2/agents-rag/)
- [Hardware & IoT](/data/steer-v2/hardware-iot/)
- [Interpretability & AI safety](/data/steer-v2/interpretability-ai-safety/)
- [Learning](/data/steer-v2/learning/)
- [Light-films & CUDA studio](/data/steer-v2/light-films-cuda-studio/)
- [Phone & device agents](/data/steer-v2/phone-device-agents/)
- [RL & training](/data/steer-v2/rl-training/)
- [Security & reverse engineering](/data/steer-v2/security-reverse-engineering/)
- [Tensor-Slayer & weight editing](/data/steer-v2/tensor-slayer-weight-editing/)
- [Trading & markets](/data/steer-v2/trading-markets/)
- [Video & sound studios](/data/steer-v2/video-sound-studios/)
- [Workstation & desktop setup](/data/steer-v2/workstation-desktop-setup/)

Data is a great tree with lit platforms in its branches · artist unknown (signed lower right) · the artist's.