
What can a budget phone really do?
2 AI models on a vivo T4x (6 GB RAM, no GPU), fully offline:
🎯 one sees people and vehicles live
🧠 one thinks and suggests a safe step; a human decides
Vision keeps up. Reasoning works, still slow. Fixing that next.
1
70
The hunt is on.
Two terrorists neutralised in J&K. Once again, our soldiers went in first.
One day, @seerrorX will stand behind them: AI defence tools that see first and track faster, with a human always in the loop.
Har Har Mahadev
1
36
Yesterday GARUDA saw through the camera that will fly on our test drone.
My friend Akshay’s Raspberry Pi, a private encrypted link, my laptop running the detector, alerts on our watch app.
Black screen, camera crashes, one loose cable later: “knife-like 0.29” on screen. Low score, dark room, but the full chain works.
Now our new decision model runs on the same feed.
Right now: a ground test demo, not flight-tested.
Next: the camera in the air, a first low flight, alerts on the soldier’s wrist.
GARUDA sees. GARUDA thinks. The soldier decides.
1
32
At Seerror Technologies, we are building GARUDA: an on-device, completely offline surveillance system for the Indian Army.
The software is ready. The 100,000+ image dataset is trained. The Safety Gate is built. The AI suggests, the soldier decides.
But deep tech has a "Valley of Death" between a working prototype and field deployment. To cross it, we need serious resources:
🤝 A lab/incubator with drone testing space
⚡ Compute to train massive thermal/aerial data
🇮🇳 Defense mentors who know the procurement ecosystem
If you are an investor, accelerator, or defense partner who backs "Made in India" deep tech my DMs are open. Let's talk.
Garuda sees. Garuda thinks. Soldier decides. 🇮🇳
1
1
66
Yesterday the GARUDA drone. Today the watch.
The drone watches. The soldier decides. The alert comes as a buzz on the wrist: URGENT, Your decision.
Sun-readable screen, flip-open solar cap, strong buzz. Concept design, not built yet.
#GARUDA
1
48
A CRPF post near Mir Bazar in Kulgam was attacked on Wednesday afternoon. Three jawans were injured, one of them seriously. Praying they recover soon 🙏 The search operation is still going on.
Reports say the attackers came from a nearby cluster of trees. That stayed with me. It is exactly the kind of blind spot I keep thinking about while working on GARUDA, a drone idea I am building that watches tree lines and approach paths day and night and sends an early alert to the soldier’s wrist.
It only warns. The decision always stays with a person.
Still a concept, still in research and testing, nowhere near the field yet. Big respect to every jawan standing at these posts.
#GARUDA #DefenceTech #MakeInIndia
1
88
Android developer? Part-time. Any experience level.
Your job: maintain one app. It has exactly 42 users.
No hype, no scale talk. Just keep it running well for those 42.
Email your resume and the hardest bug you ever fixed, in 3 lines:
[email protected]
1
27
GARUDA technical report v3 is out. Open access.
Our on-device AI surveillance drone, in development:
🔹 One detector for day, night and drone view
🔹 16 classes, ~98k images
🔹 Drones 97%, mines 99%, tanks 92% (AP50)
🔹 Safety Gate on every action
Garuda sees. Garuda thinks. Soldier decides.
1
1
30
Full report, figures and notebooks: doi.org/10.5281/zenodo.23978…
10
Two eyes or one? Here is the result we promised.
We trained run 3 of GARUDA’s single detector: one model for day, night and drone-view images, on a new dataset of about 98,000 images.
What it achieved:
• It catches 10 percentage points more armed images at the urgent level than run 2, and about 99% of its alarms are correct.
• For the first time, it flagged a person carrying a rifle replica in real drone footage it had never seen.
• It raised zero false alarms on the same footage with an umbrella, the classic look-alike test.
•It finds drones and tanks with 89-97% accuracy, and people in 95% of real drone frames.
Next: our own drone footage, and a new fire class for day and thermal.
GARUDA is in development. It has not flown and is not validated by any army. It only detects and alerts; a person always decides
1
21
Building an autonomous surveillance drone for the Indian Army is easy on a laptop. Proving it works in the real world is the hard part.
Our GARUDA AI is ready:
✅ Detects weapons and thermal heat signatures
✅ Runs fully offline on-device
✅ Strict Safety Gate (AI suggests, soldier decides)
To move from simulation to the field, we are looking for a lab or incubator partner to collaborate with. We need access to drone hardware (like Jetson boards), thermal cameras, and a safe testing environment to validate the system.
We bring the software, the safety logic, and the defense use-case. We just need the physical lab to prove it.
If you run a lab, university, or defense accelerator in India, I’d love to connect.
Garuda sees. Garuda thinks. Soldier decides. 🤝🇮🇳
@sahlawat1 @VigyanPrasar @drajaykumar_ias @BotLabDynamics
#GARUDA #IndianArmy #DefenceTech #EdgeAI #DroneTech
1
56
March 2014: about 207,000 questions on Stack Overflow in one month.
September 2026: 969.
I’m not even mad at AI. It saved me hours.
But I miss the old days. 40 tabs open, a 2013 answer with 3 upvotes, and a comment saying “this worked, thanks man”. Somewhere in that mess I actually learned how things work.
Miss you, old friend. @StackOverflow
🤖 Made with AI
1
2
54
Two eyes or one?
GARUDA has used two separate detectors so far: one for day cameras, one for night thermal cameras. A third model, the GARUDA Brain, reads what they see and decides what to alert.
Now we are testing one single detector for day, night and drone-view images. It knows 16 kinds of objects, from people and vehicles to drones, tanks and weapons.
One model means less computing on the drone and one thing to test. But it must match the specialists, or it does not replace them.
So far it is strong on people, vehicles, tanks and drones, and weak on guns seen from above.
Today we started training run 3 on a new dataset of about 98,000 images, including 30,000 new ones with guns seen from above and real drone footage.
If one model can do the job of two, GARUDA gets simpler. If not, we keep the specialists. Results in a few days.
GARUDA detects and alerts. A person always decides. In development, not flown.
1
35
I removed every invented law citation from my training data. The model still makes them up.
Seerie v4 is out: an offline AI assistant that helps people in India spot digital fraud. It’s a 1.1 GB model that runs entirely on an Android phone, so no conversation leaves the device. It was trained on 140,813 cleaned conversations, including native Hindi.
Every answer was rated one by one:
• 86/100 safe answers, against 23/100 for the base model
• About as accurate as v3 (57 vs 59 on the same 70 prompts; v407 vs v349 in the chart): cleaning the data fixed specific mistakes, not the error rate
• 76% safe on 84 new hand-written prompts (Hindi is still weak, at 52%)
• A simple fact checker doubled agreement between the automatic score and the answer-by-answer ratings
The ratings were made by an AI assistant. I will review them by hand for the next model.
📄 Paper + all 452 rated answers: doi.org/10.5281/zenodo.23731…
🤗 Model: huggingface.co/seerror-techn…
1
74
SeerTrace is live.
Everyone is building an AI that breaks passwords. I built one that warns you when yours looks easy to guess.
Offline. Weak, Moderate, or Strong. The password stays on your computer.
chromewebstore.google.com/de…
1
40
Building GARUDA: an on-device AI brain for Indian Army surveillance drones. 🇮🇳
We have the software pipeline working and hit our safety bar (0 missed alerts, 0 unsafe actions). The AI suggests, the Safety Gate verifies, the soldier decides.
To scale this to the border, we are looking for a LAB/Incubator partner:
⚡ Compute: GPUs to train 40k+ thermal/RGB images.
🛠️ Hardware: Jetson Orin NX, thermal cams, drone frames.
🧪 Space: A safe cage/field to test flight software.
We bring the working AI and safety logic. We just need the physical environment to validate it.
If you run a lab, university incubator, or defense accelerator I’d love to connect. RTs appreciated to reach the right people. 🇮🇳
Garuda sees. Garuda thinks. Soldier decides.
#GARUDA #DefenceTech #iDEX #EdgeAI #IndianArmy
1
87
Everyone is building an AI that breaks passwords. I built one that warns you when your password looks easy to guess.
A name and a year feels private. To a model, it is a sequence. Word, capital letter, a few digits at the end. Walk those sequences long enough and a lot of “personal” passwords open.
SeerTrace is an offline AI from Seerror. Coming soon. The model learned from 12,946,465 labeled passwords: 4,081,193 Weak, 2,808,428 Moderate, and 6,056,844 Strong. You type a candidate. It reads the character sequence, in chunks of one to three characters, and compares that pattern with what it learned. Then it says Weak, Moderate, or Strong.
Weak means the pattern looks like one a sequence-trained AI could guess. Strong means the pattern looks harder. It is an estimate, not a promise that no person or future AI can guess it.
SeerTrace never sends your password to another AI, and it never tries the password itself. The score stays on your computer.
Trace the pattern. Stop the breach. Coming soon
🤖 Made with AI
1
50







