
@Jodo_in Eng Director by day, Dad in evening, hacker at night.
India
Joined August 2009
- Tweets3K
- Following244
- Followers538
- Likes1.8K
One of the interesting challenges we've dealt with at Jodo is building a system that moves money based on data from systems we don't control.
Our collection pipeline handles scheduled auto-debits across eNACH and UPI Autopay. On peak days, a single presentation batch can exceed 60k instalments.
Over the years, we've learned that fetching the latest amount isn't enough. External systems can return incorrect data, gateway responses can get lost, and a retry can move money twice. One thing has remained clear: when correctness and completion come into conflict, correctness wins.
Wrote about how we built this pipeline, the mistakes we made in the initial version, and how our thinking evolved over the years. @Jodo_in @JodoEngineering
jodo.in/blog/engineering/on-…
If stars are beyond a deep-sea fish's perception, what is our equivalent of stars? @grok
1
63
Experience helps, but it doesn’t make us immune to cognitive biases. And many of these biases show up under surprisingly predictable conditions.
I’ve been thinking about what this means for how we structure decisions, encourage dissent, and build engineering processes.
The more I look at it, the more I think this is a systems problem, not just an individual judgment problem.
Wrote down my thoughts here: codingdash.com/post/cognitiv…
1
4
143
Ritesh Shrivastav retweeted
Smallest AI is now available on @RapidaAI.
Build and self-host voice agents on Rapida, and plug in Smallest AI for the voice layer:
→ Pulse for streaming STT
→ Lightning v3.1 & v3.1 Pro for TTS
Choose your models, voices, and speech provider directly from the Rapida dashboard, with no SDK required.
Open orchestration + production-ready speech models, now in one stack.
This is part of a larger series on building systems on top of Claude: codingdash.com/series/automa…
#Claude #AIEngineering #AgentDesign
31
I’ve written this down properly with examples and what this means in practice: codingdash.com/post/understa…
1
1
37
The more interesting bit is what compaction optimizes for. It tries to preserve flow, not exact details.
That works fine for chat. It starts breaking once you’re building anything that depends on correctness.
Once this clicks, a few things change in how you think about building:
- context is not memory
- anything important shouldn’t live only inside the conversation
- you need a separate layer for durable state
1
20
And that rewrite loses information. Not randomly. Predictably.
Which means if your system depends on conversation history as memory, it will drift over time. Not because something broke, but because it’s designed that way.
1
62
While working with Claude, I spent some time understanding how compaction works. The short version is: the system doesn’t really “remember” everything, it rewrites the conversation into something shorter.
1
104
Most teams are still treating Claude like a smarter autocomplete.
That works until it doesn't, and when it breaks, it breaks in ways that are genuinely hard to debug. Not errors. Inconsistency. Context drift. Outputs that were fine yesterday and aren't today.
I've been thinking about why some teams compound on this and most don't. Turns out it comes down to one thing they do differently from the start.
Wrote it up: codingdash.com/post/treating…
#claudecode #softwareengineering @AnthropicAI
5/ I wrote the full piece. Field report from inside the infrastructure.
The graveyard of correct ideas. The concurrency problem. The India opportunity. The pricing model that's being made obsolete.
codingdash.com/post/state-of…
4
70
4/ 10 years later the ideas we tried and shelved at Exotel are now funded product categories.
Sentiment analysis on calls. SOP to workflow automation. Voice biometrics.
The vision always existed. The capability layer just arrived.
But the infrastructure layer hasn't caught up. 95% of AI voice pilots are failing. Less than 1% of contact centres have autonomous agents in production.
1
1
75
3/ I think about that whenever someone tells me the telephony layer is a commodity to be abstracted away.
It isn't. At the moments that matter most, it is the only thing left.
1
24
2/ People stranded and in need of rescue could make a call or send an SMS. No app. No internet. No smartphone required.
That call automatically triggered a tweet putting their location in front of people who could help.
It worked. People got rescued.
1
49
1/ During the 2015 Chennai floods, our data centre went dark. What I remember most isn't the outage. It's what Vishnu and I built in response.
I started blogging on codingdash.com in 2014 as a college student. Same blog. Still running.
The name came from @codinghorror's Coding Horror, named after an icon in Steve McConnell's Code Complete, used to flag cautionary examples of bad code. That lineage felt right, and codingdash stuck. It's been my corner of the internet since.
Shipped a redesign this week. The second one since starting, the first redesign I did in 2020, inspired by @Stammy's site. This one was overdue not because something broke, but because the site had quietly fallen behind the person writing it.
Short note on what changed: codingdash.com/post/a-new-co…
6
109
Full writeup on why I built it and where it's headed: codingdash.com/post/prlens-a…
3
86
I spent a long time looking for an AI code review tool that actually fit how my team works. Tried a few. None felt complete.
Hidden costs. No control over which model runs. Guidelines that were either too generic or too rigid to customise. And when things broke, no visibility into why.
So I built one over a weekend and open sourced it.
PRLens — bring your own Claude or GPT-4o key, bring your own guidelines, runs as a GitHub Action in your own pipeline. No new vendor. No black box.
Still early, lots to build. But the foundation is there and I'd love collaborators who've felt the same frustrations.
github.com/prlens/prlens
1
6
153
Most AI-powered products hide their real cost structure. They bundle model inference into a flat subscription or per-seat price, the same way early SaaS bundled AWS costs. But there's a key difference: AI model pricing is dropping 50-200x per year and the provider landscape shifts every few months.
When your cost base is that volatile, bundling it creates problems:
- You can't pass savings to customers as models get cheaper
- You can't let customers bring their own API keys (which they increasingly want)
- You can't give enterprises the compliance visibility they'll legally need by August 2026
The question every AI builder should ask: what is my product worth independent of the model running underneath it?
If the answer is clear, decouple the burn and prove it.
codingdash.com/post/your-ai-…
2
4
96

