Nitish alodia

Dog parent 🐶
Head of AI/Engineering - Product Engineer
Currently building in AI healthcare sector, learning and implementing something new everyday.

[Nitish alodia @AlodiaNitish]

Agentic coding seems solved because it is very definitive in nature. Things like copywriting are far too though for agents to crack yet because copywriting has no clear definition of "correct" , it is infact a matter of "taste" which llms just don't have at all.
Extends to any kind of creative work.
Easy to fetch data and do analysis, much harder to be "creative"

[Nitish alodia @AlodiaNitish]

@_svs_ Matt pococks skills pretty much follow the same thing - https://t.co/OLGmTlQVhJ
You might want to check them out

[Nitish alodia @AlodiaNitish]

𝗙𝗮𝗯𝗹𝗲 𝟱 𝗵𝗮𝘀 𝗹𝗮𝘂𝗻𝗰𝗵𝗲𝗱, 𝗯𝘂𝘁 𝗶𝘁 𝗰𝗼𝗺𝗲𝘀 𝘄𝗶𝘁𝗵 𝗮 𝗵𝗲𝗳𝘁𝘆 𝗽𝗿𝗶𝗰𝗲 𝘁𝗮𝗴.
Here’s how to save up to 70% on tokens while using Fable 5:
Your costs can skyrocket when one expensive model handles everything—planning, typing, and re-reading the entire conversation with each turn.
The solution?
Split the tasks.
𝟭. 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 𝘃𝘀. 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻
Use a strong model to create the plan, outlining the approach, steps, and constraints.
Then, employ a cheaper model to execute each step.
This way, the expensive model focuses on crafting a compact plan instead of editing every line of code.
𝟮. 𝗘𝘀𝘁𝗮𝗯𝗹𝗶𝘀𝗵 𝗮 𝗥𝘂𝗹𝗲
Make this task division a rule in your CLAUDE.md/AGENTS.md file
Treat the main session as an orchestrator, not an implementer.
Offload substantive tasks—research, implementation, audits—to a sub-agent.
Specify the model for each task: the strong model for judgment-heavy work and the cheaper model for well-scoped or mechanical tasks. Always avoid leaving it on default.
Ensure each prompt is self-contained and distribute independent tasks in parallel.
𝟯. 𝗣𝗹𝗮𝗻 𝗦𝘁𝗼𝗿𝗮𝗴𝗲
The location of your plan is crucial.
Keep it in your project file, not in chat history.
The file serves as the artifact, while the transcript can be discarded.
𝟰. 𝗥𝗲𝘀𝗲𝘁 𝗜𝗻𝘁𝗲𝗻𝘁𝗶𝗼𝗻𝗮𝗹𝗹𝘆
Long sessions can degrade performance.
Regularly write the current state to the file, start a new chat, and reload the plan.
The trick is to make the split a rule, not a habit.
Put this in your project-instructions file — CLAUDE.md / AGENTS.md:
“Treat the main session as an orchestrator, never an implementer.
Offload every substantive task — research, implementation, audits — to a sub-agent.
On each dispatch, set the model explicitly: the strong model for judgment-heavy work, the cheaper model for anything well-scoped or mechanical.
Never leave it on default.
Keep each prompt self-contained, and fan out independent tasks in parallel.”
𝗔 𝗰𝗼𝘂𝗽𝗹𝗲 𝗼𝗳 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗻𝗼𝘁𝗲𝘀:
This method works best for sequential handoffs—plan, execute, reset. Using parallel agents with their own contexts can lead to increased costs.
The effectiveness of the cheap executor relies on the clarity of the plan. A vague plan results in poor output.
In summary, stop paying a premium model to think and type simultaneously.
Let it create the plan, allow a cheaper model to execute it, and prioritize the plan as your state.

[Nitish alodia @AlodiaNitish]

1. Built an agentic system to generate content for different platforms
2. Built a agentic system for analysing d2c ad pipelines using meta, ga4 and Shopify mcp
3. Built a system that is scheduled and runs in morning and gives me actionables for the day

All constantly learning and improving systems

[Nitish alodia @AlodiaNitish]

@realmadhuguru Its as the cloudflare ceo put it - only builders and sellers survive.
Building doesnt have to be only writing code, it can also be building a team of agents for automating flows within the company for user research, feedback loops etc

[Nitish alodia @AlodiaNitish]

I’ve been deeply interested in loop engineering and self-improving agents since openClaw and how to apply these to solve domain specific problems.
So I built an agentic pipeline around Meta Ads and Shopify, not just to automate marketing work, but to understand where these systems start breaking down in practice.
𝗧𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝘁𝗮𝘀𝗸𝘀 𝘀𝘂𝗰𝗵 𝗮𝘀:
1. Meta ads funnel analysis
2. Shopify conversion funnel analysis
3. Complete user view/clicks to purchase pipeline
4. Website audit
5. Campaign creation using past knowledge
6. Long-term memory created during the analysis phase, utilized in campaign creation

𝗧𝗵𝗲 𝗰𝗼𝗿𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗶𝘀 𝘀𝗶𝗺𝗽𝗹𝗲 Meta and Shopify operate as separate systems. The ad platform tracks clicks, spend, and placements, while the store monitors sessions, add-to-carts, and orders. Unfortunately, neither system integrates the data, leading to misaligned attribution.
𝗧𝗵𝗲 𝗼𝗯𝘃𝗶𝗼𝘂𝘀 𝗺𝗼𝘃𝗲 - Hand both systems to one AI agent and ask it to analyze the funnel.
𝗧𝗵𝗲 𝗶𝘀𝘀𝘂𝗲 𝘄𝗶𝘁𝗵 𝘁𝗵𝗮𝘁 - A single agent handling everything—from data pulling to computing rates, drawing conclusions, and writing memory—can make unseen trade-offs. It may compute conversion rates inaccurately, treat the ad platform's attributed sales as absolute truth (which they are not, due to over-counting), and rely on last month’s performance without validating if it was a genuine signal or mere noise.
Instead, I designed a system with separated concerns:
- 𝗙𝗼𝘂𝗿 𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀 serve as entry points: /analyze, /cro, /campaign, /memory. They sequence the work and manage checkpoints without reasoning.
𝗦𝗶𝘅 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗮𝗴𝗲𝗻𝘁𝘀 𝗲𝗮𝗰𝗵 𝗽𝗲𝗿𝗳𝗼𝗿𝗺 𝗮 𝗱𝗶𝘀𝘁𝗶𝗻𝗰𝘁 𝘁𝗮𝘀𝗸:
- One pulls raw data from both sources.
- One joins and normalizes the data.
- One reasons through the data step by step.
- One inspects the live store for conversion issues.
- One designs the next campaign as a tracked experiment.
- One writes to memory, but only after your approval.

A single Python script computes all statistics, while the LLM reads the output, ensuring the math is trustworthy by maintaining clear boundaries.

The memory layer distinguishes between what you’ve communicated to the system (your voice, positioning, benchmarks) and what it has learned (promoting insights only after confirming signals).
The system defaults to WAIT when data is insufficient for action, providing a clear threshold for reassessment.
Everything produced is a draft, requiring your launch and approval for every learning. This boundary ensures an honest feedback loop, as untraced decisions cannot be graded, and ungraded outcomes cannot become real learnings.
The full design, including the architecture diagram, the deterministic stats engine, the confidence-building memory layer, and the learning loop from campaign to playbook, is detailed in the article.
Comment if you want the repo for this
https://t.co/i8PAarQikk

[Nitish alodia @AlodiaNitish]

https://t.co/m1TGZt5lha

[Nitish alodia @AlodiaNitish]

https://t.co/m1TGZt5lha

[Nitish alodia @AlodiaNitish]

@heynavtoor This can actually serve as foundation for any kind of agentic systems.
On first principle - each agentic system is actually just deep search across some data and then surfacing the most critical insights from it in order

[Nitish alodia @AlodiaNitish]

@sumanthraman But Won't he be paid for the work and expectations? Do you mean the work isnt worth 16lakhs?

[Nitish alodia @AlodiaNitish]

Absolutely Agree. This is the core reason I always did product work with engineering but never shifted completely to product management.
In small orgs - PMs end up becoming project manager running errands for the CEO
In Big orgs - its a game of who can invent metrics that they can show accomplished and get promoted.

The real work of defining what metrics with actually matter ( based on business/product/user feedback ) - how to systematically increase it happens at very few places sadly.

[Nitish alodia @AlodiaNitish]

https://t.co/ACtfRFjCIg

[Nitish alodia @AlodiaNitish]

@sanjeed_i Whats the payscale like? And whats the expectation?

[Nitish alodia @AlodiaNitish]

Valid, but very difficult to make the choice tbh. B is always enticing but its so difficult to judge founder conviction/capability/product potential in 4-5 calls. Even if these hit, Ai landscape is changing so fast that what might look great on paper might be redundant in 6 months leading to a tiring job search again

[Nitish alodia @AlodiaNitish]

@dravishakatoch @somoclubdotcom

[Nitish alodia @AlodiaNitish]

@dravishakatoch @tarunrs runs something similar already. Might be worth checking out

[Nitish alodia @AlodiaNitish]

@AmirMushich How to direct

[Nitish alodia @AlodiaNitish]

The job descriptions are all just random right now.
I had a call with a recruiter and even after making complete agentic products in last 3 years, the recruiter kept asking if i knew aws bedrock 😅
With the skill sets required, the recruiting industry will also have to understand what each skillset means, its not just frontend, backend anymore, there is a whole spectrum for specific talents in engineering

[Nitish alodia @AlodiaNitish]

@NirantK Isnt that normal? I thought that was normal

[Nitish alodia @AlodiaNitish]

LLMs don't fail because of bad prompts.
They fail because of system design.
If you're building serious AI infrastructure — multi-agent systems, memory architectures, diagnostic layers, eval pipelines — and need someone who thinks at this layer:
I'm open to conversations.
#AIEngineering #AgentArchitecture

[Nitish alodia @AlodiaNitish]

The "AI Engineer" title has been colonized by full-stack engineer roles. The actual architecture-layer discipline - the one that determines whether AI systems work reliably in production - has 7 distinct skill clusters that are under-articulated and largely invisible in current job market discourse.

[Nitish alodia @AlodiaNitish]

System-level coherence over time
When Agent A, Agent B, your memory layer, and your eval pipeline all run simultaneously — they need to agree.
Not just functionally. Semantically.
If they contradict each other across steps, errors don't just stack — they compound. By the time you see the output, the source is 4 layers upstream.
Coherence is the glue skill nobody talks about.