Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions summary/sessions/305/attendees.adoc
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
==== Attendees

. link:https://x.com/Alpastx[Alpesh Bhagwatkar^]
. link:https://twitter.com/furtado_jaden[Jaden Furtado^]
. link:https://twitter.com/KartikSoneji_[Kartik Soneji^]
. AK
. Janvi Matani
. Kushagra Mehta
. Manan Gandhi
1 change: 1 addition & 0 deletions summary/sessions/305/config
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
catchup_config_noimage=1
44 changes: 44 additions & 0 deletions summary/sessions/305/content.adoc
Original file line number Diff line number Diff line change
@@ -0,0 +1,44 @@
Date: 12-09-2026

Duration: 1 hr 9 mins

==== Topics Discussed

* Kushagra Mehta explained a multi-agent setup that sends a prompt to analyzer, verifier or solver, and finalizer roles.
** The analyzer produces an answer plus a confidence score. If confidence is above 0.5, the next agent uses a cheaper or mid-tier model. It only makes small fixes.
** If confidence is lower, the setup sends the request to a stronger, more expensive model.
** The current design uses three tiers:
*** a local model
*** an intermediate model such as link:https://ollama.com[Ollama^]
*** paid models such as link:https://www.anthropic.com/claude[Claude^] and GPT-4
** Jaden Furtado said this is mainly a workflow. He said similar behavior could be specified with much less custom code. Kushagra agreed it is closer to a fine-tuned architecture than a novel idea.
** Kushagra said that a model that scores its own answer is unreliable. The model can hallucinate or overrate the result.
** Alpesh Bhagwatkar asked how routing would work when candidate models are close in quality, such as link:https://www.deepseek.com[DeepSeek^] and GLM. Kushagra said the current tiers are far apart. He said near-equivalent models still need more work.
** He said a conference on cognitive computing and generative AI accepted the paper. He said he would try to bring a demo next time.
* A discussion on how to report prohibited material found online took place.
** Kartik Soneji said the report depends on the case:
*** a DMCA request for copyright
*** a report to the people who would be harmed by a leak
*** a report to the host or link:https://www.cloudflare.com[Cloudflare^] when Cloudflare serves the site
** He also mentioned an earlier Cloudflare report. The site later went offline. It was unclear whether that one report caused the takedown.
* Alpesh asked whether to study Windows internals or AI-based testing in more depth.
** He said AI testing looks more open right now. His Windows background is closer to introductory link:https://www.offsec.com/courses/pen-200[OSCP^] Active Directory work. It is not kernel APIs, drivers, or fuzzing.
** He also described his final-year project as a custom AI harness for finding security issues.
* Janvi Matani asked whether cybersecurity is still a viable field.
** Alpesh said security work will stay necessary. The entry bar is higher because basic web app pentesting is now heavily automated.
** He treated cybersecurity as a broad study of systems, not a narrow specialty. That study includes websites, databases, APIs, and how AI agents tokenize, prompt, and act.
** For college study, he suggested that the syllabus overlap with current needs:
*** agentic engineering
*** harness development
*** context engineering
*** loop or graph-based control of API calls
** He named LLM safety tools as a way to stay current on issues such as prompt injection.

==== Projects Showcased

* Janvi Matani described her link:https://www.sih.gov.in[Smart India Hackathon (SIH)^] project on railway block management.
** Maintenance, electricity, and signaling teams currently send separate block requests. This makes it hard to judge criticality. It also makes it hard to keep routes useful for passengers.
** The idea is an AI optimizer in the middle. The optimizer ranks blocks and reduces manual coordination. A human stays in control.
** She also does link:https://spring.io[Spring^] CRUD work for the project. She uses map data, APIs, and scraping, including link:https://www.openstreetmap.org[OpenStreetMap^]. She said Maharashtra alone has about 370 corridors.
** Alpesh suggested that she talk to railway staff instead of relying only on AI-generated data. He offered to connect her with Ankush Kapoor. Ankush Kapoor's father works at CSTM.
** She also mentioned a separate fishermen-safety idea. The idea includes coastal hazard and cyclone alerts. It also includes a way for boats to communicate when a stretch of coast is unsafe.
Loading