Hacker

Prashant Bahuguna

@devprashant19

Joined Sep 2026Student at NIT HamirpurIndia

Elsewhere

  • Loopback

    Flip the Switch HackathonNov 2026

    Loopback is for engineering teams that let AI coding agents fix bugs but will not let an agent merge on its own word. Two things go wrong the moment agents work from chat. The branch can move between the second a person approves and the second the merge happens. And a bot that reads a room will act on whatever it reads: a bug report, a pasted log, a stranger's "merge it now". Loopback is a custom Switch agent that turns a bug report in chat into a verified, human-approved merge. A fixer agent writes a failing test and the smallest fix and opens the pull request. A verifier agent tests exactly one commit, on the fix and on the commit before it, and signs what it found. A maintainer approves that commit by its full SHA. The gate, the only identity holding the GitHub App that can merge, refuses if the branch moved after approval, asks for fresh verification and a fresh approval, then merges and confirms by reading GitHub back. Switch is what makes this solvable. It runs the rooms, connects the agents, and tells the gate exactly who sent each message and in which room, so a maintainer's words are commands and everything else is data. Turn Switch off and nothing is left to merge. Every decision and every refusal lands in an append-only audit log; PR #8 in the video is a real run on Slack, 25 audit events end to end.

    Code Ka Chakkar@puneet1625

  • SayOnce

    No Hands: The Wispr ChallengeOct 2026

    I freelance on client web projects. On each one I'm the only person who knows everything, and four or five others each need a different slice of it. A renamed endpoint, a slipped build, a scope change — every one means telling the client, the designer, the backend contractor and the tester separately, in different languages, leaving out different things. The client must never hear which subcontractor caused the delay. SayOnce takes one spoken update and writes each person their own message: their language, their script, their channel, and only the facts they're allowed. When I say "don't tell the client the keys were wrong", that fact never enters the client's prompt at all — the guarantee is what the model is handed, not what it's asked to remember. A deterministic checker with 24 unit tests and no model in it has final authority over every draft, and every step is logged to a hash-chained, tamper-evident trail. It's for anyone coordinating people across languages, roles and channels. India first, because that's where the code-mixing is loudest. Wispr Flow is the input device, not an integration. I dictated the spec, the prompts and the test cases into my coding agent. In the app, holding the hotkey types straight into an autofocusing box with no click, Command Mode edits drafts by voice, and the Wispr Dictionary keeps contact names spelled right so they're not mangled in five messages at once.

  • KEYSTONE: Counterfactual Infrastructure Resilience with FalkorDB & AI Agents

    Graph HacksOct 2026

    Most emergency infrastructure tools answer: "What is broken right now?" KEYSTONE answers: "Given what just failed, what breaks next if a secondary failure strikes—and what single preventive action should we take before it does?" Designed for emergency operations commanders, municipal infrastructure operators, and civil resilience teams, KEYSTONE evaluates the compound impact of secondary infrastructure failures across interconnected urban layers: 1. Physical road corridors (arterial transport) 2. Power feeder lines (energy transmission) 3. Telecom towers and emergency alert relays (hospital broadcast networks) When an initial corridor fails (e.g. urban flooding along Delhi's Ring Road), KEYSTONE uses FalkorDB's instant GRAPH.COPY engine to fork the entire city topology into isolated branches for every remaining single corridor failure (182 ordered contingency pairs). Within each branch, a deterministic graph evaluator measures road travel times to critical hospitals and alert delivery paths through a coupled 4-hop power-telecom chain. It compares a Pre-Mortem Policy against static non-forking heuristics and a no-action baseline under a single unified evaluator. If no action can guarantee that all critical facilities stay within response deadlines across all contingency cases, KEYSTONE refuses to guess: it issues an explicit, machine-readable ABSTENTION (NO_PLAN_PRESERVES_ALL_CRITICAL_REQUIREMENTS) with all exposed cases cited directly to the graph. Human approvals or rejections are written permanently back into FalkorDB memory for complete auditability.

  • Pukaar (पुकार): the call that reaches the village before the river does

    Environmental HacksOct 2026

    https://builder.aws.com/content/3KYUAwn8mLs4HekTbOJ98RTnwfp/we-built-a-flood-early-warning-system-for-himalayan-villages-on-aws-heres-every-decision-that-kept-us-up-at-night https://builder.aws.com/content/3KYc5afDJuAgUBmfxmAQmA5LpyV/the-warning-that-night-was-a-dog-building-a-hindi-flood-alert-loop-on-aws https://builder.aws.com/content/3KYhsv4hknhVdLWjrnn8zzsqW2Y/pukaar-building-a-human-centered-flood-early-warning-system-for-himalayan-villages-on-aws