HYDERABAD’S
DOJO
FOR AI builders.
Two days. One prototype. Engineers who'd rather ship than study.
Cohort 02 — September 2026 · exact dates announcing soon. Join the list, first to know.
CHAPTER ONE · WHY
INTELLIGENCE IS CHEAP. AGENCY IS NOT.
AI made smarts cheap. What matters now is deciding to build the thing in front of you.
That's the muscle we train. Two days in a small room in Hyderabad. You walk out with a prototype. Yours.
Public commentary on AI and building in general — not endorsements of this workshop.
CHAPTER TWO · WHAT
THE WEEKEND, in pieces.
No. 01
TWO FULL DAYS
Saturday + Sunday. Doors at 9. We build until we can't.
No. 02
ONE PROTOTYPE
Something that runs. Not a deck. Not a maybe.
No. 03
TOKENS, ON US
API keys, tools, materials — covered. Just bring a laptop.
No. 04
SMALL BATCH
Capped room. Every builder gets eyes on their build.
No. 05
TAUGHT BY A BUILDER
Led by Shahir — shipped AI products in daily use. Not theorists.
No. 06
STAY IN THE CREW
Plug into the meetups + hackathons afterwards.
CHAPTER THREE · CURRICULUM
CURRICULUM SATURDAY. build sunday.
Day 1: the toolkit, in three modules. Day 2: you build. We sit beside you.
DAY ONE · SATURDAY
// CURRICULUM09:30 — 13:00
FOUNDATIONS & THE MODEL LOOP
Model APIs, prompts, structured outputs, tool use. By lunch: a working agent with typed tools.
// ANTHROPIC · OPENAI · PYDANTIC
14:00 — 17:00
RETRIEVAL OVER YOUR DATA
Bring a corpus. Chunking, embeddings, vector store, evals on real questions.
// EMBEDDINGS · CHROMA · RAGAS
17:30 — 20:30
AGENTS THAT FINISH
Tool loops, planning, recovery from failure. A small agent that finishes a real task.
// AGENTS SDK · MCP · EVALS
DAY TWO · SUNDAY
// BUILD DAY09:30 — 19:00 · ONE PROJECT, ALL DAY
PICK A PROJECT SATURDAY NIGHT. ship it by sunday.
One long build sprint. Pair Day 1's patterns with the idea you walked in with. Instructors next to you — unblocking, reviewing.
By 18:00: UI, deploy, demo. You walk out with a live link.
WHAT YOU WALK OUT WITH
- ★A WORKING AI PROTOTYPE ON A PUBLIC URL
- ★THE REPO ON GITHUB, README YOU CAN LINK
- ★A FIVE-MINUTE DEMO RECORDING
- ★CODE REVIEW FROM INSTRUCTORS
- ★WHATSAPP ROOM WITH COHORT 02
CHAPTER FOUR · WHO ELSE
THE CREW around it.
MEETUPS
Monthly. Builders in Hyderabad show what they shipped. No slides.
HACKATHONS
48-hour weekend sprints. Real problems. Real judges.
BUILDER NET
Engineers + founders building with AI, not talking about it.
THE TEAM
The team behind High Agency.
SHAHIR M A
Founder · AI engineer & researcher
Builds and ships AI end to end. Haku, a native macOS agent that operates real apps, and Haku Terminal, a multi-agent workbench. A real-time voice engine in Rust running in production. Language models trained from scratch that run on an Apple Watch, with the research on arXiv. Speaks at MLX India meetups on designing models for Apple silicon.
LINKEDIN →SWADHEEN SHUKLA
Advisor · Microsoft Research
Principal PM at Microsoft Research. GenAI, multimodal, conversational systems. Helped build a medical AI model trained on millions of images. ICML Outstanding Paper. Multiple US patents.
LINKEDIN →CHAPTER SIX · STUFF WE SHIPPED
WE'VE DONE THE WORK. not just read about it.
Everything in the curriculum comes from things we built and run ourselves: agents, real-time voice, eval pipelines, models trained from scratch. You get the patterns that survived production, and the ones that didn't.
AN AGENT THAT RUNS A MAC
Haku: a native macOS agent that operates real apps through their own interfaces. An on-device model reads hundreds of screen elements and hands the agent the seven that matter. Shipped, v1.2.
// How to make an agent act reliably inside real software.
MULTI-AGENT OPERATIONS
Haku Terminal: a workbench where Claude Code, Codex and other agents converse, delegate, share memory and search past work over one protocol. 30 tools, in daily use.
// How teams of agents actually get coordinated.
VOICE AGENTS IN PRODUCTION
A real-time voice engine in Rust: WebRTC, barge-in, a from-scratch end-of-turn model that answers in 2 to 5 ms. Serves live speaking practice today. Plus a bridge that lets an AI agent answer WhatsApp Business calls.
// Latency, interruption and turn-taking, solved the hard way.
LLM PIPELINES WITH QA
The Smart Language and Supertutora: a learning platform in nine languages. LLM content generation with automatic quality checks, planner and evaluator agents, and simulated users that run whole journeys against the API before real ones do.
// Evals and quality gates for LLM output at scale.
MODELS TRAINED FROM SCRATCH
A 203M-parameter mixture-of-experts pretrained natively in ternary. Runs on the Neural Engine of an Apple Watch. The measurement study behind it is on arXiv (2608.22110, Aug 2026).
// What actually runs on-device, measured, not guessed.
PUBLISHED EXPERIMENTS
Ten write-ups at lab.haku.sh on small models and speech on Apple silicon, each stating its method and its limits. Talks at MLX India meetups in Hyderabad and Bengaluru.
// Method and limitations, every time.
FINAL CHAPTER · NOTIFY ME
DROP YOUR NUMBER. we'll ping once.
Cohort 02 opens September 2026. Exact dates, early-bird, address — you'll be first to know. Then quiet.