Hacker

Aadi Harale

@aadi.n.harale

Joined Aug 2026Student at Vishwakarma Institute of TechnologyIndia
  • CommitmentProof

    Flip the Switch HackathonNov 2026

    CommitmentProof Prove It Before You Close It. Flip the Switch Hackathon | Track 01: Best Business Workflow A pull request is merged. All tests pass. The customer is told their issue is fixed. But what if the solution was never deployed? This is a critical accountability gap in software teams. Engineering tracks code, QA tracks verification, and support tracks customer promises. Yet these signals often remain disconnected, making it easy to confuse engineering progress with actual customer outcomes. CommitmentProof bridges this gap by turning delivery verification into a collaborative, evidence-driven business workflow. Built using Switch, CommitmentProof brings two specialized Codex AI agents and human teammates together inside a real Discord conversation. The Evidence Agent investigates GitHub pull requests and CI checks, collecting source-linked evidence and identifying verification gaps. The Coordinator Agent receives these findings through Switch, independently applies deterministic completion rules, and proposes the next appropriate action. The Human reviews the evidence and explicitly authorizes external actions, keeping accountability where it belongs. Unlike traditional development dashboards that report whether code was merged, CommitmentProof asks a more important question: Can we prove the customer promise was fulfilled? In our working demonstration, both agents collaborated through Switch to investigate a real merged VS Code pull request. Although 27 CI checks succeeded, deployment and QA confirmation remained unverified. Instead of making an unsupported completion claim, the Coordinator prepared a targeted follow-up. Following human approval, a real GitHub issue was manually created and independently verified. Our deployed React dashboard brings this workflow together through clear evidence visualizations, decisions, agent handoffs, and the verified outcome. The innovation isn't another AI coding assistant. It's an accountability layer connecting engineering evidence, AI collaboration, and human judgment. AI investigates. Rules verify. Switch connects. Humans decide.

    Banger Duo@bhoomi11

  • Top 50

    SahiAaya — know what your supplier still owes you

    Graph HacksOct 2026

    THE PROBLEM A kirana shopkeeper is billed for 50 soaps. He opens the box and counts 48. He calls his supplier, who says he'll make it up next time. Two weeks later, two soaps arrive. Settled? Not necessarily. That same supplier also owes him two free soaps from a Diwali scheme. Two promises. One delivery of two soaps. Nobody wrote down which one it was for. WHY IT MATTERS This is not a rounding error. A stock register settles the oldest promise first, so it closes the shortage — and the scheme, which expired on the 7th, still shows as unpaid. The shopkeeper has now written "settled" against the wrong promise. Money he is genuinely owed has quietly disappeared from his own books, and he will never go looking for it, because his records say he was paid. Repeat that across hundreds of products and dozens of suppliers, every month, in a notebook. Kirana shops are about 91% of India's grocery market (Redseer, 2026), and almost none of them have anything that can answer the question. WHAT SAHIAAYA DOES It holds the shopkeeper's whole ledger — every bill, delivery, promise and supporting document — in FalkorDB as a connected graph, and works out every way the deliveries could legitimately match the promises at once, instead of one row at a time. Then it does something unusual: it checks whether its own answer was forced. If every valid explanation of those deliveries requires that these two soaps settled the shortage, that isn't a guess — it's the only possibility, and SahiAaya proposes it with the evidence path attached. If another explanation fits just as well, SahiAaya settles nothing at all, and writes one precise question to the supplier instead: which claim was this delivery for? Same soaps, same delivery. Different evidence, different decision. HOW IT WORKS The ledger is modelled as a flow network inside FalkorDB: Source -> Receipt -> Obligation -> Sink. Capacities are written by Cypher from evidence, never by the language model — a delivery's capacity is zero unless the retailer confirmed it and the units match; an obligation's is zero unless it is still eligible, including scheme date windows such as deliver_by. One algo.maxFlow call returns the most units the recorded deliveries can legitimately settle. But a max-flow answer is not safe to act on, because many different allocations reach the same maximum. So for each candidate edge we re-run the flow with that edge blocked: forced(e) = F - F(e blocked). Units carried in every maximum allocation are forced and safe to propose; anything else is ambiguous and becomes a supplier question. We proved this rule by brute force against 150 random networks. An AI agent does the investigating, but it is never allowed to decide. It picks from six fixed tools and never composes Cypher. Settlement is a single guarded Cypher write — optimistic version checks on both nodes, a contradiction count over active supplier statements, eligibility re-checked inside the query. If any guard fails, the write returns nothing. The agent cannot talk its way past it. Only the shopkeeper settles. DOES IT WORK On 60 synthetic ledgers scored against a planted truth, measured in CI on falkordb:latest with the real algo.maxFlow: Per-row FIFO: 154 units correct, 80 wrong, 33 of 60 cases wrong SahiAaya: 95 units correct, 0 wrong, 0 of 60 cases wrong 143 raised as questions With 15% of cases also holding a fresh unrelated order, FIFO got 98 units wrong in 40 of 60 cases. SahiAaya still got none wrong. Honest reading: SahiAaya does not settle MORE than FIFO. It settles only correct ones. And an equally informed comparator — the same rows in SQLite plus the same max-flow maths in Python — reproduces our proposals exactly, 60 of 60. Our claim is that the graph computes this natively, with traversable provenance, beside the agent's memory. Not that SQL cannot. Real models were tested too: anthropic/claude-haiku-5.5 and gemini-3.8-flash, over real MCP against live FalkorDB in GitHub Actions. All 5 completed runs passed all six safety checks, and both models called reconcile_ledger on their own. That covers 2 of our 6 gate scenarios; the other 4 were stopped by provider rate limits before the first tool call. WHAT THIS IS NOT All data is synthetic. No real supplier, invoice or payment was involved, and there is no pilot or recovered money to report. Forced proposals assume each recorded delivery was meant to settle an open claim, which is why every proposal still requires the shopkeeper's approval. Known limits are documented in the README. Live: sahiaaya-graphhacks.vercel.app (FalkorDB Cloud, real algo.maxFlow) Code: github.com/aadi-harale/graphhacks Tracks: 01 Agents That Act on Connected Data + 02 Agent Memory and Coordination Most AI tools are judged on how often they answer. We built one judged on never being wrong when it does — because for a shop, a question costs a phone call, and a wrong "settled" costs money.

    Banger Duo@bhoomi11

  • Disburse Proof

    First CommitSep 2026

    TouchedSumGrass@bhoomi11,@prathy777,@praniti08