A lost-child desk for crowded malls. An Urdu voice assistant that tells farmers which day to spread fertilizer. A firewall that stops AI agents from obeying instructions hidden in an email. University students built all of these in about 24 hours at Agentika 2.0, and every one of them is running live.
Agentika 2.0 was a two-day AI automation hackathon at NaSCon 2026, FAST NUCES Islamabad, on 16 and 17 October 2026. This was our second edition of Agentika (the first was hosted in 2024, a first of its kind). This time, Agentika was generously sponsored by Alternate Futures – the agent cloud made for developers, by developers. We (Grayhat) arranged the entire show, from pulling in Alternate Futures to its execution. We received an astounding amount of participation; 192 students from 22 institutions signed up in teams of two, and 74 teams shipped a working project.
The theme was Human + AI: Better Together. We didn't want another round of chatbot wrappers. Every team had to build an agent that does real work on its own and still hands the important decisions to a person. There was also one hard rule: the project had to be live on Alternate Clouds. A demo on localhost didn't count.
Agentika 2.0 in numbers
The 77% submission rate is the number we're particularly proud of – at our Wrap-up: Web Social Party Game Jam!, fewer than half the teams made it to the finish. This time, three out of four teams built something, deployed it, recorded a demo and posted about it on LinkedIn & X.
The crowd came from far beyond Islamabad. FAST NUCES sent 79 people, followed by COMSATS (24), Air University (22), NUST (17) and UET (12), with others travelling in from Lahore, Faisalabad, Multan, Sargodha, Layyah, Haripur and Nawabshah.
Day 1: a workshop, some riddles and a long night
Day 1 opened in a full hall. Ghulam Muhi Uddin from Grayhat ran a one-hour workshop to guide the participants on how to participate in a hackathon officially. He ran some live code demos, and showed participants how to deploy on Alternate Futures Cloud through both the UI, and through the MCP via the CLI (pretty neat!)
The Q&A ran for almost 40 minutes.
To break the ice, we ran an automation riddle challenge where right answers won T-shirts. By the end of it, people were talking to the organizers and to each other instead of staring at their laptops. Then the building started, and for many teams it went on through the night.
Day 2: deadlines and the final round
Day 2 began in the labs. Teams spent the morning fixing deployments, recording demo videos and filling in their submissions.
Once submissions closed, judges from Grayhat went through all 74 projects and picked the Top 10. The finalists then pitched live to the panel, with the following judging panel:
- Hayk Mavisakalyan (Cofounder, Alternate Futures)
- Saad Bazaz (CEO & Cofounder, Grayhat)
- Usama Abdullah (Business Operatons, Grayhat)
- Abdurrehman Subhani (Engineering Manager & Cofounder, Grayhat)
Hayk from Alternate Futures judging over a video call. The day ended with the results and a lot of photos.
The winners
🏆 Winner: Talaash by 404 Team Not Found
Abdul Raheem and Wania Rahman
When a child goes missing at a mall, an airport, a stadium or a festival, the first few minutes matter most. Most venues still run that search on radios and paper. Talaash turns it into a coordinated search desk.
- A parent reports the missing child from their phone, by form or voice note
- One AI agent builds the case, and a second one, running on a different model, checks it for red flags
- Guards log sightings, and the system matches them against the case and against found-child reports
- Alerts go out to guards on Discord, along with ready-to-read PA announcements
What won the judges over was the human side. No alert goes out until a security officer approves it, and only a person can mark a child as safe. That single check stops prank reports, and it stops anyone in a custody dispute from using venue security to track a child.
🥈 Runner-up: Sona AI by Team Oblivion
Syeda Aliza and Syed Rizwan
A farmer buys fertilizer, spreads it, and a day later the rain washes it away. Many smallholder farmers also have no cash at sowing time, and most farming apps are too complicated for someone with little or no formal education. Sona is built for exactly that farmer, and it speaks Urdu.
- The farmer talks to it in Urdu, and it answers in Urdu
- It checks live weather and builds a fertilizer plan for the crop and the size of the land
- An agronomist agent, a weather agent and an economist agent must agree before it suggests a day
- It can spot plant disease from a photo of a leaf
- It can order fertilizer on 0% buy-now-pay-later credit
The farmer always picks the day and confirms every order, and the dealer checks stock before anything ships.
The Top 10 finalists
These ten teams made the cut from 74 submissions.
| Team | Project | What it does |
|---|---|---|
| 404 Team Not Found | Talaash | Finds lost children in crowded venues |
| Team Oblivion | Sona AI | Urdu voice assistant for farmers |
| Hardcoders | RelayDesk | Fixes IT access and email outages |
| Code Catalyst | Agent Firewall | Stops AI agents from obeying hidden instructions |
| The Opus 5.5 | ReRoute | Finds a backup supplier when an order runs late |
| KMForge | Afterward | Handles the paperwork after a death in the family |
| Team 404 | Biztective | Finds out why an online store lost money |
| LoopTroops | AWAAZ | Helps harassment survivors get help and report safely |
| Powerpuff Girls | Power Hunt | Job applications with no made-up claims |
| AgentX | ClearGate AI | Flags suspicious invoices before payment |
Every project from Agentika 2.0
Here's every team that shipped, grouped by the problem they took on. Most of these are live right now, so click through and try them.
Safety, rights and public services
AWAAZ by LoopTroops · Top 10 finalist
Most harassment cases never get reported, because survivors fear being identified and complaints go nowhere. AWAAZ sends her live location to a trusted contact in seconds, walks her through a gentle, skippable interview, drafts a formal complaint plus an anonymised public summary, and escalates the anonymised case to the press if the authorities sit on it for 24 hours.
KinGuard by AI Crusader
Scammers call elderly parents pretending to be their child or a police officer, and AI voice cloning now makes those calls sound real. A parent describes the call by voice note, screenshot or text in English, Urdu or Roman Urdu. KinGuard scores the scam risk, tells the parent out loud to hold off, and alerts a family guardian with the evidence.
ReliefLine by Man in the Loop
In a flood or earthquake, help requests pour into WhatsApp as panicked voice notes in many languages. ReliefLine transcribes and translates them, pulls out who needs what and where, flags duplicates, pins every request on a live map and offers the task to the nearest volunteer.
Disaster Response Coordinator by Neural Nexus
Rescue teams get repeated, vague, multilingual field reports and have to guess which area needs help first. This n8n workflow organises the reports with Gemini, checks them against flood-area data, ranks zones by need and writes a short brief for each one.
JobShield AI by NeuralFlow
Fake overseas job offers drain people's savings through upfront visa fees and passport requests over WhatsApp. Paste an offer or upload the PDF and JobShield flags the risky claims, explains why, and hands back a verification checklist in English or Urdu.
Entitle by Git Happens
Scholarships, grants and benefits go unclaimed because the rules sit on long official pages that keep changing. Entitle tells you what you may qualify for, and every match quotes the exact line from the official eligibility page, checked word for word in code.
Witness by Glue Sticks
Fully automated benefits screening leaves wrongly denied applicants with nobody accountable. Witness extracts facts from case files with verbatim citations, runs eligibility through a seven-clause rules engine written in code, and logs every step in a tamper-evident ledger.
WageClaim Assist by AuraFex
Workers who are owed pay usually have their proof scattered across chats, invoices and screenshots. WageClaim Assist organises the evidence, masks personal details and drafts a clear wage claim letter.c
LawFlow by Ctrl Alt Elite
Over a third of people summoned to court don't show up, often because the notice is jargon nobody can read. LawFlow turns a police FIR into a proposed hearing slot and a plain-language summons in English and Urdu, then tracks delivery and flags non-compliance after 72 hours.
TenderSentry
Public tender packages run to 50+ pages written by different teams, and contradictions in dates and amounts can cost money or hide corruption. TenderSentry cross-checks the PDFs, explains each conflict in plain language and shows the clashing excerpts side by side.
Raabta Auto by T-Square
Vehicle owners forget token tax, insurance and registration renewals until the penalties arrive. Raabta Auto drafts personalised reminders from existing government records and sends them by email and SMS.
Healthcare
MedBridge by Team Beta
Patients with several conditions see a different doctor for each, and nobody sees the full medicine list. MedBridge reads prescription photos, maps brand names to ingredients and flags duplicates, clashes and old prescriptions that were never stopped.
MedPassport by ForgeFlow
Common cold and pain medicines are banned or need permits in some countries, including transit stops, and travellers find out at customs. MedPassport reads your medicine list from a photo, checks it against every stop on your trip and drafts a customs letter.
TeachBack by 2thePoint
Patients sign surgery and trial consent forms they don't fully understand. TeachBack explains each point in plain language, asks the patient to explain it back, re-explains when they miss something and flags signs of pressure or distress.
CareLoop AI by 9x9 LDF
Small clinics lose patients to missed appointments, and chronic-care patients quietly drop out of treatment. CareLoop spots the no-shows, drafts outreach messages and finds open slots, while blocking any clinical question from the AI.
MediSense AI by Agent-Verse
Appointments, records, lab reports and scans usually live in separate systems. MediSense brings them into one platform with role-based access and AI that highlights possible findings on X-ray, MRI and CT images.
Everyday life
Afterward by KMForge · Top 10 finalist
Settling a loved one's affairs takes about 13 months on average. Afterward takes the family's wishes once through Telegram, contacts every bank, telecom and utility, follows up for months and checks each "done" against what the family actually approved.
Tandem by Inferix
Tandem calls businesses on your behalf to book a dentist, a hotel or a group dinner, and pulls you into the live call with a one-tap card when a decision is yours.
Dwellio by Team Avatar
Dwellio puts tenant repair reports, manager approvals and technician visits into one trackable flow. Managers approve repairs, and the case closes only when the tenant confirms the fix.
AI safety, security and DevOps
Agent Firewall by Code Catalyst · Top 10 finalist
AI agents with access to email and files can't tell their owner's instructions from a sentence an attacker hid in an email. Agent Firewall sits in front of the tool server as an MCP proxy, scores every tool call from 0 to 100 and holds anything risky, showing the reviewer the hidden sentence highlighted.
Mandate by Forge
Nobody lets an agent spend money or delete data unsupervised, but approving everything by hand burns people out. Mandate pauses agents before risky actions, routes them to a shared inbox, and after three similar decisions proposes a plain-English rule to automate the next one. It plugs into Python, n8n, Claude Code and Cursor.
PatchGuard by Feral Flows
Dependabot opens patch PRs without proving the bug was exploitable or that the patch fixes it. PatchGuard watches CVE feeds, tests whether the vulnerability is real, attacks the patch again to confirm the fix, then sends the result to Telegram.
RelayDesk by Hardcoders · Top 10 finalist
RelayDesk is an IT service desk where an identity specialist, a network and mail specialist and a runbook specialist investigate an outage together, then propose a scoped repair. The engineer authorises the repair plan, approves exact before and after changes, and confirms recovery.
IncidentPilot by Techies
Three cooperating agents investigate phishing reports and suspicious activity for small organisations, using read-only evidence tools across Gmail, GitHub and Windows Defender. An admin approves each response action, with a reason, before it runs.
Incident-To-Fix DevOps Agent by Tech Knights
This agent detects broken pipelines and proposes a fix. If its confidence is below 0.7, it only raises the alarm and doesn't touch anything. Every fix needs a human approval through Discord.
Autonomous Incident Responder by Sigma Air
This multi-agent workflow picks up production incidents from Sentry, investigates logs and deploy history through a custom MCP server, and opens a GitHub pull request with a proposed fix. A senior engineer reviews, tests and merges the PR.
BugProof by Promptly
BugProof turns a vague customer bug report into reproducible API evidence, a GitHub issue and a reply to the customer. When facts are missing, it asks instead of inventing them.
CoJudge by The Coalition
CoJudge was built for this very hackathon. It checks deployments, repos and demo videos, flags leaked keys and late commits, and only reveals its own score estimate after the judge has scored.
Small business and e-commerce
Biztective by Team 404 · Top 10 finalist
Dashboards tell a small online store that margins fell, not why. Biztective runs a root-cause investigation over the store's own database, computes every number in SQL so each claim traces back to rows, and writes a recovery plan to Google Sheets.
Bizzii by YinYang
Four agents (sales, operations, project management and marketing) run the back office of a 5 to 20 person service business. A Risk-Trust Gate decides, action by action, whether an agent may act alone, must ask, or is refused.
ConnectXeo by ConnectXeo
ConnectXeo is an AI store operator for Shopify merchants. A coordinator routes each instruction to one of seven specialist agents for products, inventory, orders, refunds and support, then checks the result.
ShopNod by Prompt & Pray
Bakeries and corner shops lose money in quiet hours and on stock that's about to expire. ShopNod drafts discount offers within the owner's limits and sends them as QR codes to nearby shoppers.
LessHelpless by do adad deluluz
LessHelpless turns a plain-English request like "cancel the stale pending orders" into a sorted plan, then cancels and double-checks each order after approval. Nothing changes until someone clicks a single-use Approve, and paid, disputed or VIP orders are decided individually.
ReStock AI by Chat GPTwo
ReStock AI is a purchasing officer for small retailers that decides reorder quantities with Q-learning, runs routine orders and escalates expensive or unfamiliar ones with a plain-English reason.
StockSense by Agent X
Small restaurants either run out on busy days or over-order. StockSense works out what to buy from recent sales and stock levels, picks suppliers that arrive in time, and has Gemini review the plan without letting it touch the numbers.
SourceDesk by Uzashi
Five agents act as a sourcing team for solo sellers, scouting wholesale products, comparing suppliers and working out margins in code. The seller picks the products and suppliers and accepts the final plan.
MarketRadar by blazingturtles
MarketRadar researches a product idea across Amazon listings and reviews, Google demand and supplier listings, then writes a cited opportunity report and drafts a supplier inquiry. A person approves the opportunity and reviews the inquiry before any outreach.
CoolBreeze by VibeEngineers
CoolBreeze is a voice receptionist for home-service businesses that answers calls while the owner is out on a job, books routine work and logs everything in a CRM. For a gas smell or a CO alarm, it gives safety guidance and alerts the owner straight away.
SocialAgent by Tafreeh
Send SocialAgent a raw video over WhatsApp and get back an Instagram-ready Reel, with pauses cut, vertical framing, captions and music. The creator approves the edit and the caption and decides when it posts.
AutoNauts by AutoNauts
AutoNauts runs a society event end to end: registration, payment screenshot checks with Gemini vision, QR tickets, gate check-in and review links. Unclear payments go to an organiser on Telegram.
Client Onboarding and Invoice System by Paradox
This n8n system takes a freelancer's client from request form to meeting invite, agreement, invoice and welcome email. A person approves the invite, agreement and invoice through Gmail before they go out.
Freelancer Deal & Delivery Agent by Neural Ninjas
This n8n workflow takes a client's first message through clarifying questions, a priced proposal with guardrails, content creation, an AI quality check and invoiced delivery. The freelancer approves the proposal and the final work at two separate gates.
The Signal by Vakin
The Signal handles outbound prospecting for founders and small sales teams. It builds context on each prospect from your ideal customer profile and drafts the email. Every draft goes through a Slack approval gate before it's sent.
AutoVision Studio by AutoVision Studio
AutoVision Studio generates a 28-shot image and video prompt package for a single vehicle, with a shared vehicle profile that keeps every shot looking like the same car.
Supply chain, finance and compliance
ReRoute by The Opus 5.5 · Top 10 finalist
When a critical order is going to be late, small procurement teams lose days comparing suppliers by hand. ReRoute flags at-risk orders, then a team of agents finds the cheapest supplier that meets every rule and builds a recovery plan.
ClearGate AI by AgentX · Top 10 finalist
ClearGate is an n8n agent that reads invoice images, pulls out amounts and bank details, and flags red flags like unusually high amounts and missing IBANs. Nothing is paid until a person reviews the summary, adjusts the amount if needed and approves.
Cold-chain Reroute Desk by Reasoners
When a refrigerated container's cooling fails, the cargo has hours to live. This tool estimates remaining shelf life from temperature readings, proposes diversion ports and filters out the ones that are illegal or out of reach.
DemurrageGuard by MT
Shipping lines often bill demurrage for days the contract excludes. DemurrageGuard audits the invoice against the contract, gate records and port holidays in about a minute, and drafts a dispute letter with amounts checked in code.
EthiSource by AVALON-II
EthiSource vets a supplier against news, databases and customs manifests in under 45 seconds, compares that against the supplier's own claims and scores the risk out of 100.
TraceGuard by MightyAutos
TraceGuard screens supplier data for hidden forced-labour and sanctions risk, starting with garment exporters in Pakistan and Bangladesh, and logs every decision in a hash-chained record. A named reviewer decides flagged cases with a written justification.
Escalera by Time Complexity
Escalera traces a consumer's food complaint back to the exact manufacturer batch and the retailers that received it. People verify complaints, approve inspections and escalations, and record the final safety verdict.
Cosign by Autobots
Cosign is a loan underwriting analyst that clears routine applications, asks applicants pointed questions when their numbers don't match, and leaves out protected attributes like gender and age. The model can escalate a case but never approve one on its own.
ComplAi by Dubious Duos
ComplAi reads a bank's policy draft, finds compliance gaps and ties each finding back to the regulation behind it. The compliance officer decides which changes to keep.
AmendAgent by ClaudeLimitReached
AmendAgent watches regulatory feeds from the US, EU and UK, finds the contract clauses each new rule affects and drafts amendment proposals with citations. A lawyer reviews every wording change and the owner decides what gets sent.
Ombot by Ninja Duo
Ombot triages and routes bank customer complaints, with Slack and ClickUp hooked in. Fraud, refunds and account actions are escalated to a person.
Apprentice by Yume
Apprentice turns requests from forms, email and Slack into verified tasks for the right department, and learns from the manager's corrections. High-value, urgent or suspicious requests go to a manager, and they expire after four hours with no decision.
Careers, education and hiring
Power Hunt by Powerpuff Girls · Top 10 finalist
AI makes job applications fast and dishonest. Power Hunt builds an evidence ledger from your resume, GitHub and portfolio, finds jobs, tailors your CV and labels every generated claim as Verified, Inferred or Needs you. In testing it caught 95% of planted fabrications.
JobBot by Claviculars
JobBot is a Chrome extension agent that finds matching roles, fills each application in its own tab and asks you through an inbox when it hits a missing fact or a login. You approve every final submission.
RoleLander by AFK Builders
RoleLander is a research workspace that searches live vacancies and Greenhouse and Lever boards, then judges each role against your profile with source-backed evidence and drafts an application. Nothing is auto-submitted.
asanHire by Syntax Co
asanHire is an AI hiring pipeline that builds a scoring rubric from the job post, ranks resumes with evidence, runs proctored voice interviews and writes a final report. HR makes every hiring decision.
RampUp by Sialkot Gng
New engineers lose weeks to unfamiliar codebases. RampUp coaches them through real repo tasks with hints that cite exact files and lines, and opens a mentor case with the full context when it can't safely help.
GradBeacon by SolveSmith
Students applying abroad get burned by agents selling fake admission guarantees. GradBeacon finds graduate programs, compares funding and drafts emails, preferring official university sources every time.
ECHO by 2Man
ECHO is a university help desk that answers policy questions by voice and chat, grounded only in real university rules. When it isn't sure, it hands the conversation to a staff member.
ORBIT by Agentic Rebels
ORBIT helps students find opportunities they would otherwise miss, checks eligibility and turns each one into a personal roadmap.
CareerPilot by FlowForge
CareerPilot is an n8n workflow that pulls internship listings, scores how well each one matches and tracks them in Google Sheets. The student approves or rejects each listing through a form.
APAI by Team Alpha
APAI is a personality assessment inspired by Ikigai that suggests career paths that fit your strengths, along with a development roadmap.
Research and fact-checking
ClaimBreaker by Nemegun
ClaimBreaker checks research papers against their own code, datasets and results, re-runs supported evaluations in isolated containers and exports an evidence report.
Veritas by InevitableWinners
Veritas checks a claim against a full policy document with exact citations and two AI analysts, one arguing for and one against. A person reads the cited passages and records the decision.
VeritasGuard AI by Tech by A2
VeritasGuard is a fact-checking agent that investigates claims against live web evidence and weighs what supports and contradicts each one. Users approve or flag each assessment before the report is final.
TruthTrace by Kinza Team
TruthTrace flags signs of deepfakes and manipulation in images for investigators and journalists. An investigator makes the final call on authenticity.
InsightForge by Manius
InsightForge is a research chat over your own documents that notices when retrieved passages disagree on meaning, like "bank" the institution versus a river bank. It stops and asks you which meaning to follow.
What we learned
- Making deployment mandatory. "Deploy it or it doesn't count" changed the whole event. Teams stopped polishing slides and started fixing real errors, and we ended up with 74 working projects.
- A sharp theme. "Human + AI: Better Together" pushed teams past the chatbot. Almost every project has a clear answer to "where does the person step in?"
Thank you
A big thank you to Alternate Futures for sponsoring Agentika 2.0 and putting a judge on the panel. Thanks to NaSCon 2026 and FAST NUCES Islamabad for hosting us in NaSCon's Silver Jubilee year, to our judges, and to every participant who stayed up building. Of course, last but not least, thanks to every single Grayhatter who helped run this entire event; we truly love our developer communities and love contributing to them.
Want to hire the people behind these projects, or sponsor the next one? Get in touch with Grayhat.