Deploy Analytics Agents in Six Weeks .

Data stakeholders are talking AI and data leads are lost. This 6 week program gives you practical knowledge of AI for your work. You leave with a working agent and confident about what to do about AI. Stop Feeling Lost.

The future is agentic, and it's here.

Analysts use AI to write queries faster. An agent goes further. It works round the clock, answers instantly, and multiplies how much your data actually gets used. Stakeholders expect agents now.

# Growth Aha Agent
PN
Priya Nair Head of Growth · 9:41
Why did signups dip on Friday?
Aha Agent APP 9:41
🔧 queried signups_by_source · 0.8s
Down 18% vs the 4-week Friday average. One driver: paid_social. The campaign paused at 2pm and sent ~2,400 fewer clicks.
PN
Priya Nair 9:41
Hmm, I don't follow. Can you show me?
Aha Agent APP 9:41
Sure. Friday vs a normal Friday, by source:
organic
referral
email
usual paid_social
Friday signups by source

Multiply this across hundreds of users. See the difference?

Built for senior data leaders.

This program is designed around you — senior, time-poor, already fluent in the data. Not a beginner's course on how LLMs work. Here's how that translates into what you actually learn.

01

You already know the core skill

Years of knowing what a bad number looks like — the definitions, the caveats, the sanity checks. The one new move is handing that to an agent as context.

02

A room of four senior peers

Four data leaders, not a 25-person webinar. You get unstuck in the session, the attention is on your build, and you're with people who get your world.

03

One real build, sessions 1 to 12

Not a tour of features. You build one real use case end to end — none of the side quests you'd hand a junior anyway.

"Where do I even start with AI?"

The hard part is a skill you already have.

Context is the engine powering AI agents, see it in action below. Each time you hand the agent one more thing an analyst knows, and the answer goes from confident-and-wrong to correct.

context_playground live
$ analyst asks
Which venue is best for chasing a total?
agent answers confidence: guessing
Chinnaswamy, Bengaluru. It has the highest average second-innings score (178), so it looks easiest to chase at.
no context yet"Highest scores" isn't "best for chasing." The agent is reading one average and calling it done, the way a naive prompt does.
give it context
0/3
click to hand the agent what a senior analyst would already check.
context stack
  1. empty — the agent is guessing from averages
this is a scripted taste. open the real playground →

Nothing about the model changed. Every jump in correctness came from context: sample size, segmentation, the dew rule you already knew. You're starting from years of knowing what a bad number looks like.

Four seats, all senior.

Not a 25-person webinar where you watch a speaker. Four data leaders in the room, building with the instructor at their shoulder — and with each other.

who's in the room
01 Head of Analytics
02 Analytics Manager
03 BI Manager
04 AI Analytics Lead
why four

Unblocked in the session

Stuck on your setup, your access, your data? It gets fixed there and then — not in a ticket queue between sessions.

Attention on your build

Every choice is made against your data and the questions you actually get asked, not a generic demo everyone shares.

A room that's aligned

Four people who get your world. Trade AI ideas and work through real challenges together — that's half the value.

4 seats per cohort. When they're gone, they're gone.

One build, sessions 1 to 12.

Not a tour of features. One real use case, built end to end across twelve sessions — the parts of data work that matter, none of the side quests you'd hand a junior anyway. Every week ends in something committed to your repo.

4
seats
6
weeks
12
live sessions
24
hours of build
1
shipped agent
the stack you assemble, one tool at a time
Claude
Codex
OpenAI
Python
Git
Supabase
Vercel
Streamlit
Google Cloud
Docker
Jupyter
  • A project folder that compounds: four context files instead of throwaway chats
  • Your local setup and admin rights, the #1 place people stall
  • Claude Code: the plan → build → test loop
  • Reading a data question the way an agent should
  • you ship Your repo, first commits, and a context-backed project answering one real question off your data
  • A database you control: load it, query it from Claude Code
  • Your first documented data model: grain written down, one row means one thing
  • Five test questions with known-good answers
  • Versioning the work in Git; the repo is the receipt
  • you ship A dataset loaded, first queries running, a documented model and a mini-eval file
  • From raw tables to wide, documented models, the difference senior analysts feel instantly
  • The metric layer: define active user, revenue, churn once, and reuse it in every query
  • Naming and column-level docs the agent can actually read
  • you ship 6–10 documented tables and a metric layer with the definitions written down
  • The output spec: what a good answer looks like, in a file
  • Wire the stack: model, context, agent
  • Answers resolve to your definitions, with the table, query, and definition inspectable behind every number
  • you ship An agent that answers from your model instead of guessing
  • The eval harness: put a number on correctness
  • Decision boundaries: teach it to ask instead of guessing
  • Catching regressions when you change the model or the context
  • you ship An evals/ suite and an honest accuracy report
  • Deploy on hosted infra, set permissions, hand someone else the URL
  • External testers ask real questions, and you learn from what breaks
  • The README a stakeholder or hiring manager can read
  • you finish with External testers using your agent, a public repo, and a deployed app
your-agent/ open a week →
CLAUDE.md
context/
business.md # the domain, described once
schema.md
data/seed.sql
models/ # first documented model
evals/questions.md
semantic/metrics.yml # active user, revenue, churn
semantic/models.md
agent.md # how it reasons
output_spec.md
evals/harness.py
evals/report.md # accuracy, honestly
README.md # a stakeholder can read it
Dockerfile
live at agent.yourname.ahainsights.in github.com/you/your-agent — public ↗

The repo is the receipt

the honest boundary

The whole thing runs end to end on my infrastructure, on a dataset you bring or one I give you. You leave able to repeat it on your own stack. Wiring up your company's warehouse, permissions, and stakeholders is a team engagement. That's a separate conversation.

Week titles and file names are illustrative. The shape is real: weeks 1–2 stand the repo up, 3–4 build the agent properly, 5–6 prove it's correct and ship it.

You leave with a running agent.

Six weeks in, you own an agent pointed at your data. A stakeholder can trust it because every number opens up.

agent.yourname.ahainsights.in live
asked by a stakeholder
What's 30-day retention for the users who signed up in March?
01
the data model
one wide, documented table the agent reasons over — not raw events, not a maze of joins.
fct_user_activity
user_id fk → dim_users
signup_date date
activity_date date
is_active_day bool
days_since_signup int
signup_cohort month
02
the semantic layer
the metric, defined once — so "retention" means the same thing every time it's asked.
metric: retention_30d Active at least once in the day 24–33 window after signup. "Active" = one or more sessions (metric layer, defined once).
caveat, stated: a user who returns on day 40 is not counted. This is windowed retention, not "ever came back." Ask for the latter if that's what you meant.
03
the context layer
the business glossary the agent is handed, so it resolves your words your way — not its guess.
context/glossary.md
"signed up in March" → signup_cohort = '2026-03'
"active" → one or more sessions that day, not a login
"retention" → windowed (days 24–33), not ever-returned
04
the query it ran
generated from the three layers above, then run — inspectable, not a black box.
select
  count(distinct case when
    days_since_signup between 24 and 33
    and is_active_day then user_id end)
  * 100.0 / count(distinct user_id) as retention_pct
from fct_user_activity
where signup_cohort = '2026-03';
05
evals
checks that must pass before a number is shown — what makes the answer defensible when you're not in the room.
evals/retention_30d.yml 3 / 3 cases pass
known cohort → 41.2%, matches a hand-checked figure
empty month → returns 0, never null
a month with no data yet → refuses, says so

Building it for work, on your real data? You can host it on your own infrastructure instead. Ask on the call.

One price. The whole build.

the 6-week cohort
₹55,000.
₹10k deposit holds your seat₹45k before session 1

included

  • +12 live sessions: Sat & Sun, 2 hr each — 24 hr of build time
  • +Ongoing hosting on my infra; I absorb the cost
  • +A dataset you bring, or one I give you
  • +The WhatsApp room for async support between sessions
  • +A 15-min 1:1 onboarding before session 1

you leave with

  • A working AI agent, pointed at your own data
  • Confidence using AI coding tools in your day-to-day
  • A real grasp of what tokens, MCPs and agents actually mean — you can build one, not just use it

you bring

  • A paid Claude account
  • Weekend availability, six weeks
  • A laptop with admin rights
Apply for the cohort Founding-cohort pricing. 4 seats · next batch 17 October · application-first, we talk before you pay

Aadi Jain.

Founder, AhaInsights10+ years in dataEx-lead, JoyrideB.Tech, IIT Delhi LinkedIn →GitHub →About →

Analytics engineering, BI, data modeling, product analytics. The unglamorous parts of data, done for a decade, for teams I led.

I don't teach "AI transformation." I teach data people to build one real thing end to end, then a harder one.

FAQs.

Choose a question from the list below.

ask_aadi — objections.log live
// Pick a question below — it runs against the same answers you'd get from me on a call
common objections

Still not answered? aadi@ahainsights.in — real reply, not a bot.

Apply for the cohort.

The Cohort is meant for senior data professionals (5 YOE is minimum). Leave your details and I'll get in touch with you in 24 hours.

No payment now — we talk first.