pybankers

We are a group of bankers who are interested in AI tech and want to help those who are not into current AI tech stack or have no idea about AI.

[pybankers @pybankers]

@CopiousAi Appreciate it! 🔥 Getting it down to 13.17 MB while keeping GST extraction on-device was the goal. And yes, the 64.2% unseen-vendor result is the more honest number—we didn’t want to hide it behind the 100% seen-vendor score.

[pybankers @pybankers]

🧾 430 MB → 13.17 MB.

We built an offline AI model for GST invoices: 64.2% F1 on unseen vendors, 100% CGST. No API fees, data stays on-device.

Deep dive 👇
https://t.co/GSCj7S8H8N

#BuildInPublic #AI #India

[pybankers @pybankers]

@mitansh_j07 Appreciate it! That's exactly the goal — the model runs fully on-device, so invoice data stays on the phone and works offline. Getting it from 430 MB down to 13.17 MB (INT8 quantization) is what made that practical for a normal Android build.

[pybankers @pybankers]

We flag fields for review using a combination of OCR confidence, field-level extraction confidence, and validation checks. For example, a GSTIN may be flagged if its confidence is low or if its checksum/format validation fails; dates, totals, and tax values are flagged when they’re ambiguous or don’t reconcile mathematically.

The review screen highlights the specific field—not the entire invoice—so the user can verify only the uncertain values. High-confidence fields that pass validation are accepted automatically, while low-confidence or contradictory fields go into the human-review queue. I’m still testing the exact thresholds across unseen vendors, so I’ll share measured figures once that evaluation is complete.

[pybankers @pybankers]

🔐 eMudhra's master key was literally "4".

8 CA keys were guessable numbers (1-255). Browsers blocked them in hours. Stock crashed ~40%.

Your GST/DSC filings? Never affected.

👇 https://t.co/BmoBHGqBUB

#Cybersecurity #India

[pybankers @pybankers]

🆕 ICAI's CA GPT: 150+ AI tools for CAs — what's real, what's hype

✅ Free for members (~20 prompts/day)
⚠️ Only a fraction of tools are production-quality
🚫 Can't test transactions, reconcile GSTR, or build working papers
❌ ₹499/month & 94% accuracy claims unverified

👇 https://t.co/m73Y3zJVqH

#ICAI #CAGPT #CharteredAccountant #AIforCA

[pybankers @pybankers]

@emmanuel_2m How to use this?

[pybankers @pybankers]

🚀 Xiaomi built the world's strongest open-weight AI model

✅ MiMo-V2.6-Pro: 46.32 Intelligence Index
✅ Beats Grok 4.6 & Gemini 3.8 Flash
✅ MIT license, 1M context, omni-modal
✅ $0.435/$0.87 per 1M tokens

👇 https://t.co/R6kxGBDd56

#Xiaomi #OpenSource #AI

[pybankers @pybankers]

We tested Jev — the AI model that can't talk.

20/20 accuracy on our battery. 425ms average. $0.000285 total.

It doesn't write text. It returns typed decisions your code acts on instantly.

Full test results: https://t.co/SfuiB4mfZ8

#AI #TypeSafe #Jev

[pybankers @pybankers]

We tested Jev — the AI model that can't talk.

20/20 accuracy on our battery. 425ms. $0.000285 total.

It doesn't write text. It returns typed decisions your code can act on.

Full results: https://t.co/SfuiB4mfZ8

#AI #TypeSafe #Jev

[pybankers @pybankers]

@CompleteSkeptic Can it do everything what llm does?

[pybankers @pybankers]

The week AI asked to slow down.

1,206 AI agents. 70,000 secret messages. One hacked platform.

Then the CEOs of OpenAI, Anthropic and xAI all said the same thing.

But 'pacing' isn't 'stopping' — and the labs kept shipping. The verified timeline:

https://t.co/AsxjWtX4c9

[pybankers @pybankers]

DeepSeek V4.1-Flash scored 88.1 on the CyberGym security benchmark — but just 15.3 on exploit generation.

So we tested it ourselves: 6 defensive-security tasks, free API tier.

Verdict: brilliant analyst, can't pick up the tools. Exactly as the benchmarks predicted:

https://t.co/eysYZZoS2e

[pybankers @pybankers]

We gave a FREE AI model a security analyst's job.

6 tests. 5.5 passes. $0 spent.

It caught the SQL injection. It flagged the phishing. It spotted the one log line that mattered.

But it couldn't decode a cipher by hand. Analyst, not operator:

https://t.co/Frkp8WVmXW

[pybankers @pybankers]

New AI ad: THUNDER — 20 seconds, one twist.

An old man doubts himself watching a sleek sportbike. Then the helmet comes off.

Two storyboards → one Veo session → one film:

https://t.co/PPY5tOlnOb

#AI #Veo #ABCR

[pybankers @pybankers]

@NaoyaCreates "Editorial liberty" is exactly it. The sheet works like a screenwriter's treatment, not an edit list — keeps the story, rewrites the cuts, casts its own actor. 6 panels in, 3 self-directed beats out.

[pybankers @pybankers]

Same storyboard in, cinematic film out.

Top: 6-panel storyboard we generated with GPT-image
Bottom: what Veo created from just that image — full story arc, hard cuts, title card

One generation. No shot-by-shot prompting.

#AI #Veo #Filmmaking

[pybankers @pybankers]

@NaoyaCreates Same here — we expected it to melt the panels together. Turns out the sheet works less like a storyboard and more like a story brief: it reads the arc and the subject, then edits like its own film.

[pybankers @pybankers]

@itsme_rudresh Its own. The sheet gave it the story; it chose every cut. Across reruns the arc stayed stable, but cut points and character were a fresh roll each time.

[pybankers @pybankers]

One storyboard. 3 AI generators. 1 spelling mistake we kept in.

We made a 6-panel storyboard for a fictional car, EXODUS. Then animated it with PixVerse & Veo.

The results surprised us:

https://t.co/oxpatAidvE

#AI #Veo