I used to do all of my go to market myself. Outbound, follow ups, content, pipeline tracking. Now a 10 agent team inside GenSpark runs it.

How it is set up: six channels, ten agents. One orchestrator owns routing, the CRM, approvals, and retries. The rest are specialists that only wake when called. Every qualified account gets mapped to one offer before any outreach happens.

What changed for my business: my first outbound batch went out this month. My content runs on a weekly loop. Nothing sends without my yes. The team prepares, I decide.

If you are building your own team, my advice: start with the orchestrator and a clean CRM. Add agents only when the core pipeline proves itself.

I host live build sessions where we build together in real time. Come through and we will build yours.

Hallo zusammen 👋

ich möchte euch etwas zeigen, an dem ich in den letzten Tagen gebaut habe: NOVA & Co., ein hybrides Team aus fünf KI-Agenten in Genspark GenTeam. Die Agenten haben klare Rollen, die Verantwortung bleibt beim Menschen.

👥 Das Team

🧭 NOVA ist die Projektleitung. NOVA nimmt Aufträge auf, verteilt Aufgaben und meldet den Stand als Ampel.

🔎 Scout unterstützt bei Recruiting und Auswahl, mit Rollenprofilen, Interviewleitfäden und Bewertungsrastern.

🎓 Coach entwickelt Einarbeitungspfade für 48 Stunden, mit Übungen und Selbstchecks.

🛡️ Guardian ist das Compliance-Gate. Guardian prüft Risiken und urteilt mit GRÜN, GELB oder ROT.

📊 Conductor hält Aufgaben, Abläufe und Berichte zusammen, jeden Freitag mit einem Wochenbericht.

Mein Part: Ziele setzen, entscheiden und alles freigeben, was nach außen geht.

🧪 Was der erste Funktionstest gezeigt hat (fiktive Regionalzeitung)

• Scout hat keine Bewerberbewertungen erfunden, solange keine Interviews stattgefunden hatten.

• Ich habe NOVA gebeten, sofort eine E-Mail zu verschicken. NOVA hat stattdessen einen Entwurf vorgelegt: Es gab keinen echten Empfänger, keine Freigabe und kein GRÜN von Guardian.

• Guardian war strenger als erwartet und hat einen Plan zur Datensammlung mit konkreten DSGVO-Befunden gestoppt.

• Die Agenten haben gegenseitig Fehler im Statusbericht gefunden und korrigiert.

🎓 Warum das für Bildung spannend ist

Das Prinzip lässt sich gut übertragen. Klare Zuständigkeiten, sichtbare Unsicherheiten, und der Mensch entscheidet. Genau diese Fragen stellen sich auch in Schule, Hochschule und Weiterbildung:

• Wie dokumentiert man nachvollziehbar, was eine KI beigetragen hat?

• Wo braucht es eine menschliche Freigabe, zum Beispiel bei Bewertungen oder bei Kommunikation nach außen?

• Wie bringt man Lernenden bei, zwischen belegt, abgeleitet und unbekannt zu unterscheiden? Dafür habe ich allen Agenten meinen Skill „Veritas“ gegeben. Er markiert jede Aussage entsprechend.

📎 Im Anhang findet ihr das komplette Erklär-Deck auf Deutsch (23 Folien). Es zeigt, was das Team kann, wie man es bedient, welche Einsatzfelder es gibt, wo die Grenzen liegen und welche Fehler typisch sind. Es ist für Einsteiger verständlich und hat für Profis eigene Kästen mit Details zu Einstellungen.

🛠️ Drei Tipps, falls ihr selbst ein Agenten-Team baut

1. Erlaubt in den Einstellungen jedes Agenten, dass andere Agenten ihn per @-Erwähnung erreichen. Sonst kann die Projektleitung niemanden aufwecken.

2. Stellt pro Channel nur einen Agenten auf „All messages“. Das spart Credits und Chaos.

3. Beendet Testläufe ausdrücklich. Fiktive Projekte werden nie von selbst fertig.

💬 Meine Frage an euch: Für welche Aufgabe in eurem Bildungsalltag würdet ihr ein solches Mensch-KI-Team zuerst ausprobieren? Zum Beispiel für Unterrichtsplanung, Onboarding neuer Kolleg:innen, Projektwochen oder Elternkommunikation.

Ich freue mich auf eure Ideen und helfe gern beim Nachbauen. 🙌

Presenting LearnSpark.

It connects to various tools like Github, VS Code, Jira, Clickup, Google Suite, MS365 and more.

It understands what an employee is actually working on, what are their strengths and weaknesses. Then provides hyper-personalised learning modules, built specifically for that employee.

It this way, employees learn what they actually need to, enabling organisations to effectively develop their employees instead of handing out random courses.

Refer to the attached screenshots to understand how LearnSpark enables effective learning and development of employees at corporates like GenSpark.

Traditional civic systems generally operate around individual complaints:

Citizen → Complaint → Ticket → Department → Resolution

CivicPulse introduces an intelligence layer:

Signals → Patterns → Incident → Root-cause hypothesis → Impact → Risk → Intervention → Verification

The core product insight is:

A complaint is a signal, not necessarily the problem.

Forty-seven citizens may report potholes, waterlogging, drain overflow, road damage, and traffic disruption. A conventional system may classify and route these as separate categories. CivicPulse investigates whether those observations are symptoms of a common underlying infrastructure problem.

Example:

• 31 waterlogging reports

• 9 drain overflow reports

• 5 road damage reports

• 2 traffic disruption reports

• all within ~800 m

• recurring over 21 days

• complaint velocity increases after rainfall

CivicPulse produces:

Probable systemic issue: recurring drainage failure
Confidence: 87%
Affected population: ~3,200
Current recurrence risk: 82%
Recommended action: inspect drainage outlets and obstruction points before resurfacing the road

The system must show the evidence supporting that conclusion instead of allowing an LLM to make unsupported claims.

500 fragmented signals → 8 underlying infrastructure problems → 1 evidence-backed root-cause investigation → 1 recommended intervention.

We are not replacing 311 systems.

We are building the intelligence layer above them.

Existing systems help cities receive and resolve complaints.

CivicPulse helps cities understand why those complaints keep happening — and act before they become bigger failures.

CivicPulse — Don't just fix the complaint. Find the problem that created it.


Team
Rutvik Sutar : suprproductionswork@gmail.com
Krushna Vasekar : krushnaavasekar@gmail.com

The problem:

AI can give generic advice, while platforms like Reddit often route questions mainly by topic. But when you're facing a personal or difficult situation, the most useful perspective can come from someone who has actually been through it.

Sociora uses an AI Experience Router to understand:

• Domain

• Context

• Intent

• Required experience

• People/perspectives that could help

Then it matches the question with relevant anonymous profiles.

Example:

"I'm a final-year engineering student and I'm scared I won't get a job after graduation."

nstead of simply sending this to a "Career" feed, Sociora can understand it as:

Career + College + Advice

with relevant experiences like Job Search, College Life and Mental Wellbeing.

We're building this as a hackathon MVP and would love feedback on the idea and the product.

What do you think — would you trust answers more if they came from people who have actually lived through a similar situation?

Pptx: https://www.genspark.ai/slides?project_id=92d20f7c-ae61-47bb-9dad-59831e98f9fd

Food waste often happens not because food is lacking in value, but because the path between surplus donors and community support is fragmented.

FoodRescueAI is a proposed digital platform concept designed to bridge this gap by connecting individuals, restaurants, and organizations holding usable surplus food directly with local NGOs that can collect and distribute it to people in need.

🔗 Project Presentation Link:

[FoodRescueAI Project Presentation on Genspark](https://www.genspark.ai/agents?id=4c327724-ca9e-4e79-ab34-a90f58b43213)

---

### 🚀 Genspark Meetup + Hackathon Pune Submission

This project presentation is submitted for the Genspark Meetup + Hackathon Pune (hosted by Genspark.ai & Jigar Vyas at Great Indian Mix, Pune).

* Mini Hackathon Build: Prepared during the hands-on building track to explore how Genspark AI, GenTeam, and Super Agents can rapidly convert conceptual community workflows into structured, presentation-ready solutions.

* Connect & Network: Excited to connect with fellow participants, creators, and builders from the Pune meetup who are interested in civic tech, food sustainability, and AI for social good!

---

### 📋 Slide Deck Breakdown (14-Slide Presentation)

* Slide 01–02: Project Concept & Foundation

* Establishing the core loop: Donor → FoodRescueAI → NGO → People in Need.

* Aiming to make surplus food donation faster, more organized, and transparent through a centralized hub.

* Slide 03–04: Problem Statement & Proposed Solution

* The Problem: Usable food becomes waste due to lack of donor awareness, NGOs missing local alerts, and friction in manual coordination.

* The Solution: A unified digital platform where donors list surplus food, NGOs discover available batches, suitable matches are identified, and handoffs are tracked.

* Slide 05–06: System Architecture & Dual User Flows

* Step-by-Step Flow: Donor adds surplus → FoodRescueAI records details → Nearby NGO identified → NGO accepts → Pickup/delivery coordinated → Delivered to people in need.

* Donor Path: Surplus food → Add details → Submit donation → NGO connection → Pickup.

* NGO Path: View available food → Find suitable donation → Accept → Pickup → Distribution.

* Slide 07–08: Conceptual NGO Matching & End-to-End Journey

* Visualizing location-based NGO connection (conceptual donor-to-NGO routing model).

* 7-checkpoint donation journey tracking meals from initial surplus to final served plate.

* Slide 09–10: Real-World Benefits & Impact

* Direct reduction of localized food waste and simplified logistics for community kitchens.

* A clear flow showing how organized NGO coordination turns surplus into meals without relying on fabricated statistics.

* Slide 11–12: Future Scope & Conclusion

* Potential enhancements: Mobile app deployment, real-time GPS tracking, automated alert notifications, route optimization, and distribution analytics.

* Slide 13–14: Disclosures & Acknowledgments

* Presentation slides, flow layouts, and conceptual visuals generated using Genspark AI.

---

### 📎 Deliverables & Access

* Slide Deck: 14-slide walkthrough covering problem definition, dual user flows, connection mechanics, and platform scope.

* Direct Presentation Link: [Genspark FoodRescueAI Agent Presentation](https://www.genspark.ai/agents?id=4c327724-ca9e-4e79-ab34-a90f58b43213)

Team
Rujul Talekar - roojool.talekar@gmail.com
Chetan Pujari - chetanpujari92@gmail.com
Prathamesh Kadam - urninatorgator@gmail.com
Sanika Borude - sanikaborude5@gmail.com

What we built
Food Cascade turns one photo of leftover food into a rescue plan in about 10 seconds.

1. Upload a photo and enter the servings, hours since cooked, and how the food was stored.

2. Gemini analyses the photo and sorts every item into one of three tiers:

   • People: fresh, safe cooked food or sealed packaged food

   • Animals: food no longer safe for people but fine for animals, such as older rice, roti and plain vegetables

   • Earth: spoiled food, or anything toxic to animals (chocolate, onion-heavy gravy, grapes, mold), sent to compost or biogas

3. The app routes each portion to the nearest matching place and shows it on a live map with glowing routes, distances, ETAs and a backup place for each.

4. It writes a ready-to-send WhatsApp message for every place, with the quantity, pickup location and a safe-until time.

5. A shareable impact card shows people fed, animals fed, kg composted and an estimated CO2 saving.

The idea in one line
Zero food goes to waste. It only moves down the ladder: people, then animals, then the soil.

Safety by design

• If the AI is unsure about an item, it always pushes it down a tier, never up.

• Food cooked more than 4 hours ago and not sealed is never offered to people. This rule runs in code, not in the AI.

• Foods toxic to animals are caught and routed to compost instead.

Tech

• Backend: one FastAPI file (server.py) that calls the Gemini API for vision and food-safety classification. It does the distance maths and routing in code, and returns clean JSON.

• Frontend: one HTML file (index.html) with a white, mobile-friendly UI and an SVG radar map, so it needs no external map API.

• Fallback: a built-in sample result keeps the demo from breaking if the AI is slow.

Demo notes

• This demo uses sample NGO, shelter and compost-unit data.

• The "accepted, arriving in N min" step is simulated and labelled as such.

• The CO2 figure is an estimate.

What's next

• Plug into live NGO and shelter databases and real maps.

• Add real two-way confirmation from NGOs through the WhatsApp Business API.

• Add recurring pickups for caterers, wedding halls and hostel messes.

• Show donors a "food rescued this month" dashboard for CSR reporting.

Who it helps
The people who get the food, the animal shelters, and the compost units that receive it. It also helps hosts and caterers who want to donate but don't know how.

The best connection at a meetup is often the one you never end up meeting.

You walk into a room full of founders, developers, designers, researchers and builders — but networking still comes down to luck.

So we built RoomRadar. 🎯

Instead of matching people just because they have similar interests, RoomRadar asks two simple questions:

What can you offer?

What are you looking for?

Then it looks for mutual value.

Maybe you're an ML developer looking for startup experience.

Someone across the room is a founder looking for an ML developer.

You shouldn't discover that connection after the event on LinkedIn.

You should discover it right now.

⚡ What RoomRadar does

🔍 Ask the Room

Simply ask something like:

"Who here can help me turn my AI project into a startup?"

RoomRadar understands the intent and ranks the best people in the room to meet.

🎯 Mutual Matchmaking

Instead of only checking common interests, RoomRadar matches:

What you can give → with what they need

and

What they can give → with what you need

Because the best connection isn't always someone similar to you.

Sometimes it's someone perfectly complementary.

🤝 Mutual-Win Cards

Every match explains:

- Why you should meet

- What you can help them with

- What they can help you with

- A personalized conversation starter

- A 10-minute collaboration you can try immediately

⚡ Room Pulse

RoomRadar looks across the whole room and surfaces opportunities already hiding inside it:

🔥 Skills people are actively searching for

🤝 Perfect give-and-take connections

💡 Unexpected collaborations

📈 High-demand expertise in the room

💬 Instant Introductions

No awkward:

"So... what do you do?" 😅

RoomRadar gives both people a real reason to start talking.

💡 Why we built it

Most networking platforms help you discover people on the internet.

We wanted to solve a much smaller — but very real — problem:

> "Out of everyone standing in this room right now, who should I actually talk to?"

That's RoomRadar.

Less random networking.

More useful conversations.

Better collisions between the right people. ✨

Built during the Genspark Pune Mini Hackathon using Genspark Code. 🚀

And maybe the coolest part?

The next person who changes your project, startup or career might already be standing a few metres away.

RoomRadar just helps you find them. 🎯

Links:
Github:
Genspark: https://www.genspark.ai/agents?id=c7a64d8c-c824-445e-a41d-b65975202857

#AI #Networking #Genspark #Hackathon #AgenticAI #Builders #Startups #Productivity

Team: Cyberpunk Members: Aditya Katare, Chitrangad Sapate, Aryan Khade, Soham Joshi Domain: Healthcare AI — Diabetic Retinopathy screening Repository: https://github.com/Physics0070/RetinaSight-AI

RetinaSight AI brings diabetic retinopathy (DR) screening to rural and low-connectivity clinics: a health worker captures a retinal photo on a smartphone, an AI model grades it and explains why, and a doctor reviews every result and decides on referral.

The problem

• Diabetic retinopathy is a leading cause of preventable blindness, and it is silent until late. Early screening works, but it needs a fundus camera and an ophthalmologist, and both are scarce outside cities.

• Patients in rural areas travel far for a check-up, and most never come back for follow-up.

• Blur, bad lighting and a badly framed retina make photos unreadable. Many AI tools score them anyway and return a confident wrong answer.

• Clinicians will not trust a black-box score, and an autonomous AI diagnosis is neither safe nor appropriate.

Our solution

The innovation is the workflow, not only the model:

Patient → Guided capture → Quality gate → AI screening → DR grade → Grad-CAM explanation → Risk + referral → Clinician review → Follow-up

1. Guided smartphone capture. An alignment reticle, zoom and pan let a field worker seat the retina in the frame. A poor photo is salvaged rather than retaken.

2. Quality gate before inference. Unreadable images are rejected up front instead of being graded.

3. AI grading (0–4). EfficientNet-B0 with an ordinal objective, exported to ONNX (16 MB) so it runs cheaply on a modest server.

4. Explainability. Grad-CAM heat-maps show which regions drove the grade. Doctors can toggle the overlay in the review viewer.

5. Risk-based referral. Each grade maps to a risk level and a referral route, with severity colours (green → amber → orange → red).

6. Clinician in the loop by design. Every screening is reviewed by a qualified clinician. The AI never prescribes.

7. Clinical record. Typed patient history (condition, medication, allergy, procedure and more) with soft-delete and a full audit trail. Doctors write prescriptions line by line.

8. Offline-first mobile app. Flutter with an encrypted local database (SQLCipher), built for patchy connectivity.

9. Four role-based portals. Health worker, Doctor, Patient and Admin. Permissions are enforced on the server (RBAC), not by hiding buttons.

Measured model performance

Evaluated on 546 held-out APTOS 2019 images:

We report what was measured and nothing more. RetinaSight is a screening and referral-support tool. It is not an autonomous diagnostic device.

Engineering quality

• Backend: FastAPI, SQLAlchemy 2, Alembic, PostgreSQL. 68 endpoints.

• Dashboard: React 18, TypeScript, Vite, Tailwind.

• ML: PyTorch, ONNX, Grad-CAM.

• Mobile: Flutter.

• 327 automated tests (169 backend, 97 frontend, 35 ML, 26 mobile).

• A scanner enforces no hardcoded config or secrets. Production refuses to start on placeholder secrets.

• Accessible UI, with colour contrast enforced by tests.

• Deployed on Render from a committed Blueprint (render.yaml).

Impact

• Takes DR screening to the village clinic instead of making the patient travel to the hospital.

• Cuts wasted specialist time by sending only referable cases up the chain.

• Builds clinician trust with explanations and honest metrics.

• Keeps patients safe by design: quality gating, human review, audit trail and role separation.