Hacker

Prashant thakur

@prashantthakur77

Joined Sep 2026Student at NIT HAMIRPURIndia
  • Clearance

    Flip the Switch HackathonNov 2026

    In Switch today, an agent's approval card goes to whoever may address the agent, and by default that's the person who asked. In our test setup, a persona asked for a ₹40,000 refund, approved every card herself, and the money went out. Clearance is a Switch agent plus a small gate in front of business APIs. The gate holds the credentials, so an agent can only ask, and Clearance holds the call until an eligible person answers. The person who asked can't approve. The card goes to someone else by role and rota, with an escalation timer behind it. It leads with fields parsed from the real call, and the agent's own sentence comes last, labelled unverified. It works wherever your team already talks. Every card, refusal, escalation and answer goes through Switch, so the same flow reaches a team on Slack, Discord, Teams or Mattermost through Switch's bridges, with no change to Clearance. Our demo and measurements ran on Mattermost. Approval gates decay, so Clearance also tests attention: practice requests with planted flaws, typed confirmation on high-risk cards, and autonomy for small refunds that an agent earns from a record two people sign, lost after one incident. It works for two people as well as a company: a builder and someone they trust, over their own agents and money. No LLM sits in the decision path. Identity comes from Switch's own events, never from an agent's text. If anything is missing, the answer is hold. Measured with real agents: with no Clearance, Claude's agent went ahead with risky actions in 5 of 20 runs and OpenCode's in 13 of 18. With Clearance, 0. All 25 cases in our attack suite were prevented, also on real Switch and Mattermost. Every number comes from a script in eval/.

    DellFargo@sohamjuneja

  • Debrief

    No Hands: The Wispr ChallengeOct 2026

    Developers lose the end of every day writing about work they've already done: the commit message, the pull request description, the issue comment, the standup, the client update. By then they're tired, so they write "fixed stuff" and the why is gone by morning. Debrief is for every developer who writes about their work after doing it. You talk for sixty seconds with Wispr Flow, in English or Hinglish. Debrief reads your real work (commits, diffs, CI runs, the pull request, issues, your why notes), writes every update for the person who reads it, and links each sentence to its evidence. A claim your work doesn't back is flagged before anyone sees it. You fix it by voice, approve, and it ships to GitHub or opens in the app you send from. Wispr Flow is in every step. I dictate the debrief, edit drafts with Command Mode, and my Scratchpad devlog and Notetaker meetings come in as evidence through Wispr's MCP. Debrief also exports my code's names to Flow's Dictionary so identifiers come out right. I built it by voice too: almost every commit in the repository was written by Debrief from why notes spoken into Flow.

  • Faultline: an on-call agent that can't blame a deploy until FalkorDB proves it

    Graph HacksOct 2026

    It's 2 a.m. and checkout is failing. Three changes shipped in the last hour. The on-call engineer has to pick one to roll back, and if they pick the wrong one, the outage gets worse. AI agents don't help much here today, because they answer with confidence instead of proof. We built 140 incident worlds designed to fool them, and a strong LLM working alone blamed an innocent change 37 times. Faultline is an on-call agent for SRE and platform teams with one rule: no path, no claim. It may not blame a change until FalkorDB shows two things. First, there is a causal path from that change to the symptom (algo.SPpaths). Second, that path has enough capacity to carry the damage we actually measured (algo.maxFlow). If either check fails, the agent says "I don't know" and hands the incident to a human. We think that honesty is the most important feature. When the proof holds, it does what a good on-call engineer would do: - finds the smallest containment with a max-flow min-cut, and tries it first on a GRAPH.COPY of the graph - works out who should fix it (HarmonicCentrality) and pages them - takes one human-approved action: a real GitHub pull request, or a data repair it proposes from its own measurements - checks that the action really landed through a second channel: GitHub plus telemetry, or a warehouse rebuild with its invariant suite - writes the decision back to the graph, so the next incident starts warm What we measured: - 0 confident wrong blames across 1,540 generated incident worlds, and 0 across 40 windows on the real OpenTelemetry Demo - with Claude driving the investigation: 40 of 40 right causes and 0 wrong blames The same graph also blocks risky pull requests before they merge. It answers "who owns this, who knows it, and who decided it" with node ids you can open and check.

    CoffeeAndCode@vaishali3837,@puneet1625,@lakshaydeol

  • Build It

    Beacon Night Shift: the on-call agent that fixes the 3 AM page with your voice

    First CommitSep 2026

    https://builder.aws.com/post/3Jb5v7ouXDReILJ7WMuHrlB1leL_p/why-transcript-is-the-safety-artifactvoice-approved-aws-remediation-with-strands-and-step-functions