ai-workflows on S Anand

AI tax returns 2026

Tue, 04 Aug 2026 05:48:31 +0800

On 16 July, my auditor sent me a draft Indian tax return: a refund of Rs 3 lakhs.

I gave ChatGPT my AIS, Form 26AS, bank statements, mutual fund statements, property papers, travel records, prior returns, and so on, told it not to look at the auditor’s draft, and asked it to calculate my tax independently.

It calculated a refund of about Rs 2.8 lakhs, roughly Rs 20K lower. (Less money, but a smaller refund felt less worse than a larger tax.)

Why? The auditor treated a Rs 40K in Form 26AS as interest on an income-tax refund. ChatGPT said (based on an earlier tax intimation) that it was tax deducted on an interest of Rs 1.3 lakhs.

My auditor disagreed, so we got on a screen-sharing call. She showed me Form 26AS. I showed her the earlier tax intimation. (It was confusing.)

“I think it was a mistake. I’ll rectify it,” she said.

I had a large capital gain from redeeming Indian mutual funds. My auditor said it’s taxable in India.

ChatGPT found a similar case, Anushka Sanjay Shah v. ITO. A Singapore tax resident had redeemed Indian mutual funds. The court said that the gain was taxable only in Singapore under the India-Singapore tax treaty.

My auditor said “No, only NRE accounts are exempt, not NRO accounts”. ChatGPT said “But the judgement does not mention the type of account at all”, which I told my auditor.

She checked with a senior colleague. The next morning she called. “You are eligible”, she said, “I will file it based on that scenario.” (I asked for that in writing.)

Yes, AI got me a better (or more correct) return. But frankly, it could have gone either way (been wrong, gotten me a worse return).

The main lesson I’m taking away is: The auditor cross-checked AI and filed it as their responsibility.

The useful Expert + AI workflow, for me, is:

Let AI calculate independently without the expert.
Give it raw sources, not summaries. (It can crunch lots more data.)
Ask the expert precise questions with evidence.
Get the expert’s position in writing for important things.
Let the expert take the final call.

That division worked rather well, I think.

IIM Alumni AI Workflows Workshop

Sun, 21 Jun 2026 13:01:47 +0800

The theme of yesterday’s workshop for the IIM Alumni at Singapore was Tools and Workflows was:

Agents are getting smarter, so they know what to do.
Tools agents can use are growing and are more powerful.
This combinatorial explosion creates explosive possibilites.

This workshop covered the following six workflows:

Leverage transcripts. Use Google AI Studio to transcribe non-sensitive recordings with a reusable “don’t miss anything” prompt. AI Studio’s record button is a ready-to-use transcriber.
Simplify dense text as a comic, an infographic, a story. Image generation is now a tool call an agent runs for you. Then compress it as AVIF on Squoosh before you email it to a thousand people.
Verify - cheaply. Paste one suffix: “Break this into key claims, mark certainty, flag the five highest-risk ones, and tell me how to verify or falsify each.” Convert to a skill to automate. Cross-checking with multiple models took error from 14% to 0.7%.
Skills are assets. A skill tells the agent “here’s how I do stuff.” Build them slowly, edit them weekly, and they compound for years. No skills support in your tool? Keep them as copy-pasteable prompts.
Brainstorm by forcing range. Ban the five obvious ideas; borrow from unrelated domains; smash two random concepts together with the Ideator. Hallucination is a feature when you’re being creative.
Schedule tasks. Weekly regulatory scans, daily meeting prep, market briefings - and even an “unreasonable gesture” nudge. As AI hides the tech, human relationships gain value.

Here’s the talk video and full story + transcript.

Repurposing blog posts for talks

Sun, 22 Feb 2026 22:09:49 +0800

Recently, I’ve re-used my own writing / transcripts as context to LLMs. For example, I’ve used:

This repurposing can be used for so many things.

For example, before delivering a talk to journalists “Review my Feb 2026 LLM posts and generate a single-sentence, ELI15 high-impact use case for journalists.” gets me list of use cases. Now, all I have to do is show what I did and share how it’s relevant for them, like:

I found old friends with Gemini Deep Research. You can trace sources who changed names.
I transcribed the entire Dilbert archive. You can transcribe scanned court records for $20.
… and so on.

I can do this for any field.

The approach is simple. I copy recent blog posts related to LLMs via:

rg -l "^[[:space:]]*- llms" -g "*.md" | xargs rg "^date:" | sort -k2 -r | head -n 30 | cut -d: -f1 | uniq | xargs uvx files-to-prompt --cxml | xclip -selection clipboard

… and ask the LLM:

Pick articles with the best application to [FIELD]. For each, write a single-sentence, ELI15 use case DIRECTLY following from the article. Focus on high-impact, common scenarios solving a specific problem (a persona trying to do X but blocked by Y). Skip articles without a clear, practical application. Be ruthlessly concise: ~40 words per use case. Format each line as: Article Title: Use case description.

Here are a few examples (and they’re good ones):

See journalist use cases

Extracting AI Advice: A journalist who has obtained hundreds of FOI documents or deposition transcripts can use a cheap large-context model to pull one-sentence bullets from each and then ask a stronger model to rank the top patterns – turning months of reading into an afternoon.
Finding Old Friends with Gemini: An investigative journalist can use Gemini Deep Research to trace sources who have changed names, employers, or countries – surfacing government nomination lists and public records that bridge who someone used to be with who they are now.
Gemini 3 Flash OCRs Dilbert Accurately: A journalist who receives a dump of scanned physical documents – leaked government files, court records – can make the entire archive full-text searchable for roughly $20, without waiting weeks for manual transcription.
Organizing PDF Receipts: An investigative journalist who obtains a dump of expense or procurement PDFs can ask an AI coding tool to parse each one and rename it to a standard format – making it trivial to spot duplicate amounts, unusual vendors, or suspicious date gaps at a glance.
RIP, Data Engineers: A data journalist who obtains the SQL query logs of a public institution can use an AI agent to cluster what questions the database was actually built to answer – revealing whether an agency is doing tactical monitoring or genuine public-service analysis, which can itself be the story.
Transcript AI-ded Interviews: A journalist who has done twenty interviews for a long-form profile can feed all the transcripts to a large-context model and ask it to surface the best quotes on a given angle, or flag contradictions between what different sources said – without re-reading every word.
Using AI for Work News: A beat journalist without a research desk can set up a Google Workspace automation that scans their sources weekly and delivers a single Gemini-written brief – so nothing from a regulator’s filing or a council report slips through unread.
Writing Articles from My Blog Posts: A reporter with years of stories on the same beat can feed their archive to an AI, ask it to identify which past pieces form the strongest foundation for a new investigation, and get a draft synthesis written in their own voice – rather than starting from a blank page.

See civic use cases

Breaking Rules in the Age of AI: A government adult literacy programme can drop its AI ban and replace it with instant AI-graded feedback – letting learners ask questions in their own language, fail and retry freely, and delegate tedious steps – re-engaging adults who had already given up on education.
Extracting AI Advice: A city welfare department can transcribe its 300 community consultations into one-sentence bullets and rank the top 10 concerns across meetings in the residents’ words – for a few dollars and hours.
Finding Old Friends with Gemini: A social welfare officer can use Gemini Deep Research to trace former programme participants who have moved or changed names, surfacing public records that link old and new identities – closing the loop on services that were started but never finished.
Gemini 3 Flash OCRs Dilbert Accurately: A city archivist can make thousands of scanned land records, court orders, and petitions fully searchable by running them through Gemini Flash at roughly $20 for an entire archive – no in-person visit required.
Organizing PDF Receipts: An NGO finance coordinator can ask an AI coding tool to write a script that reads every vendor PDF, extracts the date, amount, and reference, and renames the file to a standard format – turning a half-day filing chore into a five-minute step before uploading to a compliance portal.
RIP, Data Engineers: A government department can expose its data culture – and fix silent metric misalignment between teams – by feeding its SQL query logs to an AI agent that clusters them and proposes a small set of shared standard tables, no data warehouse project required.
TDS Comic Generation: A district health office with no design budget can generate multilingual public health comics for low-literacy communities by defining a few recurring characters, writing their dialogue in plain language, and letting Gemini produce each strip in minutes – at near-zero cost per language.
Transcript AI-ded Interviews: A government communications officer can feed all auto-generated meeting transcripts on a topic to a large-context model and get a press-ready 150-word statement grounded in what the department actually said – in an hour instead of a week of document hunting.
Using AI for Work News: A district administrator receiving eight separate departmental reports can set up a 20-minute Google Workspace automation that delivers one weekly email surfacing cross-department clashes – like a public works delay that will knock out a scheduled health camp – that no individual report would flag.
Writing Articles from My Blog Posts: A civic think tank analyst with 40 research reports can ask AI to pick the strongest op-ed angle for a target publication and draft it from their own words – in an afternoon, not a week of rewriting from scratch.

See community builder use cases

Extracting AI Advice: A community manager who has years of recorded Q&A sessions and AMAs can map-reduce all the transcripts to find the top recurring questions members actually ask – then build a self-serve knowledge base from members’ own words rather than guessing what to put in it.
Finding Old Friends with Gemini: A community builder trying to re-engage lapsed members can use Gemini Deep Research to find where they are now – new employer, new city, new name – and reach out at the right career moment rather than to a dead email address.
Using AI for Work News: A community manager can set up a weekly automation that scans public sources for what members have been doing – new articles, talks, job changes, launches – and auto-drafts a “members in the news” section for the newsletter that would otherwise go unwritten for lack of time.
Transcript AI-ded Interviews: A community builder who runs office hours or mentorship sessions can synthesise all the session transcripts to surface the top recurring problems members raise – then use that signal to design better programming rather than guessing what the community actually needs.
TDS Comic Generation: A community builder can create a recurring comic strip with a few consistent mascots to announce events, explain community norms, or celebrate member milestones – something memorable and shareable in a way that a plain-text post is not, and producible in minutes with no design budget.

When I delivered the Society for Clinical Data Management keynote, they audience was surprised how much I knew about their field because I spoke about Informed Consent Forms and Extracting Schedule of Assessments and so on. Truth is, I know nothing about these. Claude created the slides. I asked it to explain enough so I can talk through it.

I didn’t get the implication then, but I think I do now, and the implication is stunning. I now have material to deliver a talk to any audience.

So far, I’ve been limiting myself to technical talks. Why bother? I can speak to any audience about using AI in their field.

Human Resources & Organizational Design: HR leaders are drowning in qualitative data (interviews, performance reviews, employee sentiment) and are terrified of being left behind by AI.
Marketing & Communications: CMOs are under pressure to produce more content with fewer resources. They want to see live workflows of how a single blog post or chat transcript can be repurposed into a full campaign.
Finance & Banking: This sector is heavily regulated and drowning in unstructured paperwork. Extract specific clauses from 100-page compliance documents will immediately capture the attention of risk officers.
Event Management (MICE): The industry that organizes conferences is itself looking to modernize. Matching attendees, transcribing massive archives of past events, or predict logistical needs is highly relevant.

Time to get out of my comfort zone!

When LLM prices fall 10x every year

Fri, 20 Feb 2026 09:25:45 +0800

In Feb 2024, Claude 3 Opus was the best model, at $15/MTok.
In Jul 2024, GPT 4o Mini reached that quality at 10% of the price.
In Dec 2024, DeepSeek v3 reached that quality at 1% of the price.

Video

See the interactive version

If the price continues to fall 10x every 11-12 months or so (and it has been), then in a year, a Claude 4.6 Opus like model will cost 1/10th of the $5/MTok today, and in 2 years, 1/100th of that.

(We’ll be using better models, of course.)

But 2 years isn’t far away. If Opus 4.6 were 100x cheaper, I could do 100x of what I could do with it today. What would we do with it?

If we assume that they’ll become 100x faster as well (and that’s an important assumption), and the reliability will continue to improve, then:

LLM LSPs. Language servers could be LLMs. Hover over a squiggly line to understand a bug it spotted, right click and fix. Move on.
LLM pre-commit hooks. Write docs, write and run tests, refactor - automatically before you commit.
Continous refactoring. LLMs auto-refactor the code, run tests, and commit better code.
Auto-fix from logs. Log analysis -> Test case -> Fix -> Deployment can be automated.
Pick best option. LLMs generate 30 diverse options for each task, test all, and pick the best. E.g. What’s the best language / framework for this? What’s the better visual design? What should I build?
Live docs. LLMs auto-update docs every commit.
Adversarial workflows. LLMs continously run adversarial test cases to break the code, and fix it.
Build & discard, don’t buy. Most tools are easier to create than purchase. They’re also easier to throw away. To hell with code quality!

Using AI for work news

Sat, 14 Feb 2026 14:32:16 +0800

This week, Namit and I met a Straive team that operates from a client office. One team member asked:

I believe that we are doing wonders out here, but we are closed from what is happening in the rest our organization.

I want team members to interact with others to see what interesting things they have delivered and where we can implement that solution.

Could we have sessions, maybe a monthly newsletter, showing what innovations we’re working on? This would really keep us engaged with the tech that is going outside of the work that we do.

A good point. This reminded me of an experiment last month.

Google Workspace Studio lets you create automations. For example, here’s a flow I set up to create a weekly newsletter about client news:

Step 1: On a schedule. Every Monday at 8am…
Step 2: Ask Gemini. Scan my Google Workspace for the latest client related news and organize it as an email newsletter. Also find the latest news about these clients from the web and put it together seamlessly. Write it in a nice Malcolm Gladwell style narrative.
Step 3: Ask Gemini. Just write today’s date in YYYY-MM-DD format. Nothing else
Step 4: Draft an email.
- To: me.
- Subject: Weekly news [Step 3: Content created by Gemini]
- Message: [Step 2: Content created by Gemini]
Step 5: Create a doc.
- New doc name: Weekly news [Step 3: Content created by Gemini]
- Content to add: [Step 2: Content created by Gemini]
- Location for new doc: Weekly news

Now, I get to see a nicely formatted newsletter every Monday morning about new client activity, added to a Weekly news Google Drive folder accessible to Straive.

For example, I learnt that:

One client blocked coding agents (Codex, Claude Code, etc) which has slowed down our team.
Another client needs to modify their API Gateway + Lambda infrastructure to handle long-running agents without exceeding API timeouts
Yet others requested omni-channel convergence optimization, price sensitivity modeling, payments infrastructure setup, and much more.

I always wished we had an “internal reporter” at Gramener, going around, interviewing teams, and writing interesting stories about what’s happening.

I did not expect I’d be able to hire Malcolm Gladwell (or anyone I want) as our internal reporter!

How to Organize Browser Workspaces with LLMs and Data

Mon, 07 Apr 2025 04:44:36 +0000

Here’s an example of how I am using LLMs to solve a day-to-day workflow problem.

Every day, I interact with a barrage of websites: emails, news, social media, and work tools across multiple devices. Microsoft Edge’s workspaces syncs groups of websites across devices. I’ve never tried it, started today, and wondered: how should I organize my workspaces?

Rather than think (thinking is outdated), I used LLMs.

Extract Browsing History

Edge stores website history in a SQLite database. But the file is locked by the browser by default. So I spent a fair bit of time figure out how to read it despite it being unlocked. Here are some options:

datasette .config/microsoft-edge/Default/History --nolock
sqlite3 'file:.config/microsoft-edge/Default/History?mode=ro&nolock=1' 'SELECT url FROM urls' > urls.txt

(DuckDB cannot read locked SQLite files - else I’d use that.)

Then comes extracting the hostnames from the URLs. I used llm cmd to ask Gemini 2.5 Pro:

llm cmd 'Extract just the hostnames from urls.txt which has a list of URLs, one per line. Only pick the https:// URLs. Save into hostnames.txt'

I expanded the response awk -F/ '/^https:\/\//{print $3}' urls.txt into:

awk -F/ '/^https:\/\//{print $3}' urls.txt | sort | uniq -c | sort -k 1n > hostnames.txt

That gave me ~1,400 hostnames.

Cluster with LLMs

I passed these to O1 Pro and Gemini 2.5 Pro:

Here are the sites I visit, with rough frequency. On Microsoft Edge, I can create workspaces. Based on this browsing behavior, what kinds of workspaces might I create? Give me multiple options.

Both gave a similar set of strategies, which I’ve implemented as:

Main: email, calendar, tasks, etc.
Work: work related sites (drive, expenses, HR platform, etc.)
Chill: YouTube, Minesweeper, Netflix, etc.
Read: blogs, articles, stuff I need to catch-up on
Code: GitHub, StackOverflow, CodePen, etc.
Chores: government services, shopping, etc.
AI: ChatGPT, Gemini, Perplexity, etc.

I was surprised how similar a strategy both models converted to. Either these models really think alike, or my browsing pattern is a fairly common one. (My guess is the latter.)

Write with LLMs

After setting up my groups, I needed to write this post. Instead of slow typing, I stepped out and talked with ChatGPT. (Talking to a machine in the office felt strange, so I changed my space.) I explained my whole process, and in about eight minutes, the first draft was done. Normally, writing takes much longer, but the voice chat made it quick and smooth.

The editing after that was manual and took 20 minutes.

Things I learnt

Simple Patterns: My browsing history shows clear patterns. AI helped me find groups I couldn’t see before
Small Fixes - Big Wins: A small challenge (opening a locked file) taught me a bunch of new useful stuff
Voice Made It Easy: Talking with ChatGPT made writing fast and easy. It shows that speaking to a machine can save time

Things I Learned - 10 Nov 2024

Sun, 10 Nov 2024 00:00:00 +0000

This week, I learned:

OpenFreeMap is a free embeddable OpenStreetMap tile server. You can use MapLibre GL (more features) or Leaflet (simpler) to render it. It offers styling and self-hosting.
Zapier Actions are an easy way to set up custom actions like GMail / Google Calendar APIs for GPTs, since GPTs’ callback URLs keep changing. But they fail often, and don’t work on mobile. At least for me.
LLM Vision Use Cases in manufacturing and earth sciences (via Shivku)
- Automated geoscience image descriptions Ref
- Interpret Wind Turbine photos and charts, construction monitoring, equipment maintenance & charts Ref
- Forecast weather based on cloud photos! Ref
- Analyze thermal image of solar panels, electroluminescence images for warranty claims, ROI estimates from Google Sunroof rooftop images Ref
- Corrosion detection in electricity towers, turbines, storage tanks, penstock. Interpret non-destructive test images Ref
Google counts auto-completion when saying “25% of all the code is written by AI at Google”. “It’s a helpful productivity tool but it’s not doing any engineering at all. It’s probably about as good, maybe slightly worse, than Copilot.” YCombinator
Workflow for AI video creation: Use Meshcapade (meshcapade.com) to generate body movement of a 3D-rendered character. Pass that video to Runway’s video-to-video model to generate any visual. Add music from Suno Ref
Someone sorted the X and Y columns independently for regression. Ref
Android keyboard learning only sends model changes back to server and not local keywords. Model changes are aggregated! Ref
Here is a prompt for audio transcription using Gemini. Ref
- Transcription: Accurately transcribe the audio clip in the original language. Include all spoken words, fillers, slang, colloquialisms, and any code-switching instances. Pay attention to dialects and regional variations common among immigrant communities. Do your best to capture the speech accurately, and flag any unintelligible portions with [inaudible].
- Translation: Translate the transcription into English. Preserve the original meaning, context, idiomatic expressions, and cultural references. Ensure that nuances and subtleties are accurately conveyed.
- Capture Vocal Nuances: Note vocal cues such as tone, pitch, pacing, emphasis, and emotional expressions that may influence the message. These cues are critical for understanding intent and potential impact.
Here are some approaches to large-scale classification of medical codes. ChatGPT
- Fine-Tuning LLMs on Medical Data: Enhance LLMs by training them on medical datasets, such as clinical notes and discharge summaries, to improve their understanding of medical terminology and context.
- Multi-Agent Frameworks: Implement a multi-agent system that simulates real-world coding processes with distinct roles (e.g., patient, physician, coder, reviewer, adjuster). Each agent utilizes an LLM to perform specific functions, enhancing interpretability and reliability. ArXiv
- Retrieve-Rank Systems: Develop a two-stage system where the LLM first retrieves potential ICD-10 codes and then ranks them based on relevance, improving precision in code assignment. ArXiv
- Embedding-Based Approaches: Use LLMs to generate embeddings for ICD-10 codes and medical texts, facilitating the matching of texts to appropriate codes through similarity measures. GitHub
- Hierarchical Classification: Leverage the hierarchical structure of ICD-10 codes by first classifying texts into broader categories before assigning specific codes, reducing complexity and improving accuracy. ArXiv
- Two-Stage Verification Models: Combine LLMs with verification models, such as Long Short-Term Memory (LSTM) networks, to validate and refine the codes suggested by the LLM, balancing recall and precision. ArXiv
- Also, a mixture of models approach might work. Feed any existing NLP model / rules as a second opinion.
GraphRAG is better if data is naturally graph-structured. Else, it’s slow and fills up the context window with even vaguely related stuff. Vigneshbabu, AMAT.
ChatGPT for Windows desktop supports real-time voice and a global shortcut (Alt Space).
uithub converts GitHub repos to Markdown. Just replace “g” in “github.com/…” with “u”. Example
WebContainers are a thing and Bolt.new uses them!
Docling by IBM converts PDF, DOCX, etc. to Markdown. Like PyMuPDF4LLM but better.
Check out Loom and Cleanshot are the recommended tools for screen recording and screenshotting. But Loom is paid and Cleanshot is Mac only.
The Rubik’s cube has a Hamiltonian cycle through every one of its 43 quintillion states. Ref
OmniParser is great at parsing screenshots and identifying bounding boxes.
Recraft.ai is currently SOTA in text to image. It’s fairly impressive and could be a good alternative to Figma.
Zed.dev is an AI code editor by the creators of Atom. It’s written in Rust and is blazing fast. It has native AI integration.
Artificial Analysis has a bunch of new leaderboards and arenas.
- Open AI TTS leads the TTS Leaderboard. ElevenLabs is a bit behind.
- Recraft V3 > Flux 1.1 leads Text to Image Leaderboard
Hertz-Dev is an open source realtime voice chat model. But it doesn’t fit in Google Colab T4’s RAM
Chain of Thought reduces performance where thinking makes humans worse. Ref. Specifically:
- Artificial grammar learning
- Facial recognition
- Classifying data that has exceptions
Creating a LLM-as-a-Judge That Drives Business Results by Hamel Husain.
- Get THE domain expert (or approver) as the tester.
- Create a dataset that is DIVERSE.
- Covers EACH combination of:
  - Features
  - Scenarios: e.g. multiple matches, no match, ambiguous request, invalid/incomplete input, unsupported feature, system error
  - Persona: e.g. new user, expert user, non-native speaker, busy professional, technophobe, elderly user
- Generate data using existing data + synthetic data for each SPECIFIC combination of the above
- Evaluate based only on PASS/FAIL with a CRITIQUE detailed enough for a new employee. Include:
  - Nuances: Something a failed response did well or a passed response didn’t quite do well
  - Improvements: Suggest how model can improve
- Build an SPA to make it easy for the domain expert to review
LLMs can be made to unlearn (copyright material) better by identifying components related to the knowledge to unlearn and applying a larger learning rate to these while leaving other parts unchanged. As opposed to low learning rates for all components. Ref