1
Cost and caching
Why a 12-turn chat bills 43,800 input tokens, what prompt caching actually saves, and how one timestamp in the prompt quietly kills the cache.
43,800 tokens billedcache hits back ✓
Workshop
Where AI
systems break.
A 4-hour hands-on session for engineers building with LLMs. You run five small AI systems, watch each one fail, and write the fix.
Sat 14 Nov 2026
4 hours
Python, runs offline
no API key
online or Bangalore
No hype. No tool lists.
how AI actually works
what you build
Five labs. Five failures.
Each lab hands you a working system with one file left to write. The tests tell you when it works. The experiment after it shows you why it matters in production.
2
Retrieval with permissions
A role filter applied after search starves the results and the bot says "no policy found". You move the filter inside search, then measure recall before and after a reranker.
0 of 10 visiblefilter inside search ✓
3
Evals that mean something
Why 95% to 97% on 100 test items is noise, how many items you need before a change is real, and how one critical regression should block a prompt that scores "better".
+2% on 100 itemssignificance check ✓
4
Agent guardrails
A compromised model tries to move ₹38,000. You write the policy code that stops it without reading a word of the prompt.
₹38,000 transferblocked in code ✓
5
Memory and state
Four context tiers under one token budget. A running summary acts on a customer's stale address, and you decide what should survive into the next session.
ships to the old addressstate kept as facts ✓
RULE: you write the fix, the tests tell you it works
this workshop: no slides to watch, code to run
who it's for
A good fit if you
- Are a backend, mobile or platform engineer, or an EM, moving into AI systems
- Have shipped an LLM demo and want to know what breaks at scale
- Can read basic Python (a Kotlin/Java-to-Python cheat sheet is included)
Not a fit if you want
- An intro to ChatGPT or prompt tips
- Model training or fine-tuning from scratch
- Slides to watch rather than code to run
seats
Two ways in.
Workshop
₹3,000
- The full 4-hour session
- All five labs, yours to keep
- Live Q&A throughout
workshop + design review
₹5,000
- Everything in the workshop
- A 1:1 review of an AI system design you bring: your current project or one you're planning
Reserve a seat
Message me on LinkedIn with your tier and online or Bangalore. The format follows demand. I'll send a payment link; your seat is held once it's paid.
₹5,000 · Bangalore Message me on LinkedInfor engineering teams
Run it privately.
The same labs, run for one team, with the examples pulled toward what you're actually building. Useful when a team is moving from an LLM demo to something customers depend on, and needs a shared picture of where cost, retrieval, evals and agent actions go wrong.
- Half day, online or at your office in Bangalore
- Labs run offline on engineers' laptops; no data leaves your network
- Optional follow-up review of one of your team's designs
Tell me your team size and what you're building. I'll come back with a plan and a price.
Ask about a team sessionwho's running it
Abhishek Shukla. 13 years building mobile and backend systems, now building and teaching AI systems. Every lab here comes out of an experiment I ran for @ai.transition. More about me and the series →