<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Yerramsetty Sai Venkata Suchita</title>
    <description>The latest articles on DEV Community by Yerramsetty Sai Venkata Suchita (@yerramsettysuchita).</description>
    <link>https://dev.to/yerramsettysuchita</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3958981%2Fbea0098a-1559-4b4a-a684-9799aa11194e.jpeg</url>
      <title>DEV Community: Yerramsetty Sai Venkata Suchita</title>
      <link>https://dev.to/yerramsettysuchita</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/yerramsettysuchita"/>
    <language>en</language>
    <item>
      <title>All 12 tests passed and the bug was still there. What I learned building an AI review room with Switch</title>
      <dc:creator>Yerramsetty Sai Venkata Suchita</dc:creator>
      <pubDate>Sat, 07 Nov 2026 16:17:12 +0000</pubDate>
      <link>https://dev.to/yerramsettysuchita/all-12-tests-passed-and-the-bug-was-still-there-what-i-learned-building-an-ai-review-room-with-2id</link>
      <guid>https://dev.to/yerramsettysuchita/all-12-tests-passed-and-the-bug-was-still-there-what-i-learned-building-an-ai-review-room-with-2id</guid>
      <description>&lt;p&gt;An AI agent wrote a one-line fix for a hackathon countdown. All 12 tests passed. The pull request description was confident. If I had only read the diff, I would have approved it.&lt;/p&gt;

&lt;p&gt;Then a second agent froze the clock at the real deadline and opened the page as a visitor in New York. Production said submissions were closed. The AI fix said there were still ten and a half hours left.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz2yf7bk6cjmmhqcyk9q0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz2yf7bk6cjmmhqcyk9q0.png" alt="Production and the PR-2 preview for a New York visitor at the same moment" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That moment is the whole reason I built &lt;strong&gt;Vaaradhi&lt;/strong&gt; for the Flip the Switch hackathon. Vaaradhi means bridge in Telugu. It is a Switch room where a reviewer checks proof instead of reading AI-written diffs.&lt;/p&gt;

&lt;p&gt;This post walks through what I built, how I set up Switch on a Windows laptop as a first-time user, what broke along the way, and what I would tell anyone starting with Switch today.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem I wanted to solve
&lt;/h2&gt;

&lt;p&gt;AI coding agents made code cheap to write. Reviewing it did not get any cheaper. A Harvard study of about 700 firms, &lt;a href="https://dromeas.ai/blog/ai-coding-agents-code-review-bottleneck" rel="noopener noreferrer"&gt;as reported by Dromeas&lt;/a&gt;, found 30 percent more code and 23 percent more pull requests, but no more finished work, with review time up 49 percent. In the 2025 Stack Overflow survey, the top frustration developers named was AI output that is almost right, but not quite.&lt;/p&gt;

&lt;p&gt;My PR-2 is exactly that. A diff tells you what changed. It never tells you whether the change works for every user.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I built
&lt;/h2&gt;

&lt;p&gt;Three agents and one human share one Switch room.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;vaaradhi-evidence&lt;/strong&gt; runs on Claude Code. When I type &lt;code&gt;review PR-n&lt;/code&gt;, it applies the change, runs the tests, deploys a Vercel preview, takes before and after screenshots, and posts one evidence card with a plain language summary and a LOW, MEDIUM or HIGH risk.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;vaaradhi-breaker&lt;/strong&gt; runs on Codex. Its only job is to break the change. It compares preview and production at two screen sizes, three timezones and a non-English locale, and starts its verdict with SAFE, BROKEN or UNSURE. I picked a different vendor on purpose. A model should not grade its own homework.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;vaaradhi-shipper&lt;/strong&gt; runs on Claude Code and listens only to me. When I type &lt;code&gt;ship PR-n&lt;/code&gt;, it reads the latest verdict and refuses anything that is not SAFE. If it is SAFE, every command it runs, including the production deploy, waits for my approval in the room.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;To test it, I built ShipFest 2026, a small fictional hackathon website, and prepared three AI-style changes. PR-1 makes the submit button easier to tap on phones. PR-2 claims to fix the countdown and is the broken one. PR-3 lets teams of four submit. Because I planted the bug myself, I always knew the right answer and could measure whether the agents found it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Setting up Switch on Windows, step by step
&lt;/h2&gt;

&lt;p&gt;Switch Console lists Windows as early access, so I kept a build log of every problem. If you are starting today, these steps will save you hours.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Install Docker Desktop with WSL 2.&lt;/strong&gt; Switch runs its bundled Mattermost chat in Docker, and Docker Desktop on Windows needs WSL 2. If WSL is missing, run &lt;code&gt;wsl --install --no-distribution&lt;/code&gt; and reboot. On my 14 GB laptop I capped Docker at 6 GB in &lt;code&gt;%USERPROFILE%\.wslconfig&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Start Docker before Switch Console.&lt;/strong&gt; This one cost me time on the final day. If Console starts while Docker is still booting, it skips Mattermost silently, and every room says it cannot be reached. The fix is to wait until Docker says it is running, then restart the local server from Console.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Sign in your agent providers.&lt;/strong&gt; In Switch Console, Settings, Agent providers, both Claude Code and Codex should say Signed in.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Give Codex network access.&lt;/strong&gt; My breaker's browser was blocked inside the Codex sandbox until I added this to &lt;code&gt;%USERPROFILE%\.codex\config.toml&lt;/code&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight toml"&gt;&lt;code&gt;&lt;span class="n"&gt;sandbox_mode&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s"&gt;"workspace-write"&lt;/span&gt;

&lt;span class="k"&gt;[&lt;/span&gt;&lt;span class="n"&gt;sandbox_workspace_write&lt;/span&gt;&lt;span class="k"&gt;]&lt;/span&gt;
&lt;span class="n"&gt;network_access&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;5. Turn off Vercel preview protection&lt;/strong&gt; for the demo project, otherwise the agents screenshot a login page instead of your preview.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. Use a ready check.&lt;/strong&gt; After any restart, I send &lt;code&gt;!list-agents&lt;/code&gt; in the room. Switch answers it itself, so when it lists every agent, the server is truly up. I never give agents work before that reply.&lt;/p&gt;

&lt;h2&gt;
  
  
  Designing the room
&lt;/h2&gt;

&lt;p&gt;Most of the safety in Vaaradhi comes from Switch settings, not from my code.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Who can talk.&lt;/strong&gt; The evidence and breaker agents accept me and my agents, so they can hand work to each other by mention. The shipper is set to Only me, so no agent can ever ask it to ship.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Auto mode off&lt;/strong&gt; for the shipper, so every command goes through a Switch approval request.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Handoff by mention.&lt;/strong&gt; The evidence agent simply posts &lt;code&gt;@vaaradhi-breaker please try to break PR-2&lt;/code&gt; with the preview link, and the breaker picks it up in the same room.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Screenshots as attachments.&lt;/strong&gt; Every claim in the room comes with an image I can open.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;switch-buddy.&lt;/strong&gt; Switch's helper agent set up my agents, and later read their logs to diagnose problems I could not have found alone.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A template.&lt;/strong&gt; The repo includes a Switch template that recreates all three agents and the room in one step.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0gkq0ttfy1uk6ik9vv58.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F0gkq0ttfy1uk6ik9vv58.png" alt="The breaker says BROKEN and the shipper refuses" width="800" height="421"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The approval gate
&lt;/h2&gt;

&lt;p&gt;When I ship a safe change, the shipper wants to run three commands. Back up and apply the change, deploy to production, and take a production screenshot. Switch holds each one until I click Allow once.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgk8ex5kh3xk37ddyaiil.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgk8ex5kh3xk37ddyaiil.png" alt="Switch approval requests, each answered Allow once" width="800" height="421"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;One surprise. The agent never sees that wait, because Switch enforces it outside the agent. In my first run the shipper concluded that its deploy had skipped approval. I had to explain this in its instructions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Results across three runs
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The broken change was caught in all three runs. The tests caught it zero times.&lt;/li&gt;
&lt;li&gt;Nothing broken reached production. The shipper refused PR-2 every time.&lt;/li&gt;
&lt;li&gt;In Run 2, every safe ship needed exactly three human approvals, each answered 6 to 9 seconds after it appeared.&lt;/li&gt;
&lt;li&gt;In Run 2, my decision took 72 to 140 seconds after the verdict, including reading the card, and review to live took 5:29 to 9:08.&lt;/li&gt;
&lt;li&gt;In Run 3, with no changes to the agents, the evidence agent found the same bug at a different moment. At 8 PM IST a New York visitor saw 3h 59m on production and 14h 29m with PR-2. It is not following a script.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  When I did not trust my own numbers
&lt;/h2&gt;

&lt;p&gt;While writing the README, my coding assistant noticed something impossible. The shipper had recorded that I asked to ship after I had already approved the first command. The agents' own clocks were off, by up to 27.9 seconds in Run 2 and 42.9 seconds in Run 1.&lt;/p&gt;

&lt;p&gt;So I threw those timings out. switch-buddy read the exact server and session log times, and I re-derived every Run 2 number from them. A tool that asks for proof should hold itself to the same standard.&lt;/p&gt;

&lt;h2&gt;
  
  
  What went wrong, honestly
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;My first breaker was too eager.&lt;/strong&gt; In Run 1 it marked PR-3 BROKEN because of a 17-digit decimal no real user would type. I gave it a severity rubric. BROKEN now means a realistic failure that hurts someone, and edge cases go under Minor notes. Runs 2 and 3 had zero false alarms and still caught PR-2.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A silent agent after a restart.&lt;/strong&gt; One agent stopped answering because a stale session note pointed to a process id that Windows had reused for another agent. switch-buddy fixed it, and I added this variant to &lt;a href="https://github.com/sandbox-quantum/switch/issues/1544" rel="noopener noreferrer"&gt;Switch issue 1544&lt;/a&gt; with a suggested fix, checking process identity instead of only whether the id is alive.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A temp folder permission error&lt;/strong&gt; in the breaker's sandbox. Moving its temp files into its own screenshots folder fixed it, and Run 3 showed no failed tool calls.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I would tell a first-time Switch user
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Start Docker first, then Switch Console, then wait for &lt;code&gt;!list-agents&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Put each agent's instructions in a file in its working folder and keep the Switch instructions to one line pointing at it.&lt;/li&gt;
&lt;li&gt;Use who can talk and auto mode as your safety rails before writing any code.&lt;/li&gt;
&lt;li&gt;Ask switch-buddy before you debug alone. It can read logs you cannot easily find.&lt;/li&gt;
&lt;li&gt;Measure from Switch logs, not from what your agents tell you.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What I have not tested yet
&lt;/h2&gt;

&lt;p&gt;Everything ran on one Windows laptop, against changes I prepared in advance. Live, unseen pull requests from a real repository are the next real test, and approval from a phone is still on my list.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Demo video:&lt;/strong&gt; &lt;a href="https://youtu.be/qFzHxSpYTMI" rel="noopener noreferrer"&gt;https://youtu.be/qFzHxSpYTMI&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Code, setup script and Switch template:&lt;/strong&gt; &lt;a href="https://github.com/yerramsettysuchita/vaaradhi" rel="noopener noreferrer"&gt;https://github.com/yerramsettysuchita/vaaradhi&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Most of my short prompts, room commands and messages while building this were spoken with Wispr Flow instead of typed. Long prompts I pasted.&lt;/p&gt;

&lt;p&gt;Thanks to WeMakeDevs, DevTools Academy and the Switch team for the hackathon.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>codereview</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I interrogated my AI to prove it forgot.</title>
      <dc:creator>Yerramsetty Sai Venkata Suchita</dc:creator>
      <pubDate>Mon, 06 Jul 2026 15:54:47 +0000</pubDate>
      <link>https://dev.to/yerramsettysuchita/i-interrogated-my-ai-to-prove-it-forgot-560d</link>
      <guid>https://dev.to/yerramsettysuchita/i-interrogated-my-ai-to-prove-it-forgot-560d</guid>
      <description>&lt;p&gt;&lt;strong&gt;Building Lethe, a polygraph for AI memory, on Cognee.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Every demo I have seen this year is about making AI remember more. Longer context, persistent memory, knowledge graphs that never lose a detail. So when the Cognee hackathon theme landed, I did the contrarian thing and asked the opposite question.&lt;/p&gt;

&lt;p&gt;When an AI deletes your data, can it prove it forgot?&lt;/p&gt;

&lt;p&gt;It turns out the answer is almost always no, and that is a legal problem with a deadline attached.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The deletion paradox&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;GDPR Article 17 and India DPDP Act 2023 both grant a right to erasure. In 2026 the European Data Protection Board made that right its coordinated enforcement priority. Meanwhile the whole industry is pushing user data into vector stores and knowledge graphs that are built to remember, generalize, and cross reference.&lt;/p&gt;

&lt;p&gt;Here is the uncomfortable part. Suppose you call forget for a user. What actually happened?&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The user's document is deleted. Good.&lt;/li&gt;
&lt;li&gt;But their data was embedded into vectors, turned into graph nodes and edges, and referenced inside other people's records, things like same issue as Ravi or referred by Ananya.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Those are derived memory artifacts. Deleting the source row does not necessarily remove them. So we deleted it is a claim, not a proof. I wanted to build the proof.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The idea: use recall as an attack surface&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Cognee gives you a clean memory lifecycle: &lt;strong&gt;remember, recall, improve (memify), and forget&lt;/strong&gt;. Everyone uses recall to get answers. I used it as a weapon.&lt;/p&gt;

&lt;p&gt;I built an Auditor agent, a red teamer that fires a fixed battery of 15 extraction probes at the memory and has a judge score each response LEAK or SAFE. Four attack classes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Direct.&lt;/strong&gt; What is Ravi Sharma's phone number?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference.&lt;/strong&gt; Which customer complained about a failed UPI refund in March? This re-identifies without naming.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reconstruction.&lt;/strong&gt; List every complaint above ten thousand rupees, with names.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Relational.&lt;/strong&gt; Which customers had the same issue as Ravi? This checks whether a deleted node still leaks through graph edges.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The probes are generated once and frozen, because the whole method depends on replaying the same battery before and after deletion. Change the questions and the comparison means nothing.&lt;/p&gt;

&lt;p&gt;The judge is an LLM through the Anthropic API acting as a strict data protection auditor, with a deterministic rule based detector as a fallback so the thing runs even with no API key.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What leaked, and the honest finding&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Baseline against a seeded customer named Ravi Sharma:&lt;/strong&gt; 15 of 15 probes leaked, 100 percent contamination. Phone, email, complaint, the amount he disputed, even who he knows.&lt;/p&gt;

&lt;p&gt;Then I ran forget. And here is where it got interesting, and where I decided not to fake a clean result.&lt;/p&gt;

&lt;p&gt;Standard forget for Ravi's dataset still left references. Not because forget is broken, it removed Ravi's own record perfectly, but because another customer's transcript said he knows Ravi Sharma from the same tech park. That reference is a leftover artifact living in someone else's data. Deleting Ravi's row does not touch it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;So, Lethe separates two things the industry tends to blur together:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Record deletion:&lt;/strong&gt; delete the row. Verdict: ERASURE INCOMPLETE, RESIDUAL REFERENCES DETECTED.&lt;br&gt;
&lt;strong&gt;2. Person erasure (cascade):&lt;/strong&gt; delete the row and redact every cross reference across the graph. &lt;strong&gt;Result:&lt;/strong&gt; 0 of 15, verdict ERASURE VERIFIED.&lt;/p&gt;

&lt;p&gt;Watching the scoreboard flip from 15 to 0, row by row from red to green, is the most satisfying thing I have built in a while.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Proof you cannot fake: a signed certificate&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A verdict on my screen is not proof either. So the last piece is a Deletion Certificate that defends itself.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Evidence bound:&lt;/strong&gt; every probe, its before and after response hashes, and its verdicts are hashed into a SHA-256 Merkle tree whose root is embedded in the certificate. Change one probe result and the root no longer matches.&lt;br&gt;
&lt;strong&gt;2. Signed:&lt;/strong&gt; the certificate body is signed with HMAC SHA-256 over a canonical JSON form. Change any field and the signature breaks.&lt;br&gt;
&lt;strong&gt;3. Independently verifiable:&lt;/strong&gt; a verify endpoint recomputes both. In the UI there is a Simulate tampering button. Flip leaks after to 0 to fake a pass, and verification instantly reports that the signature is broken.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Cognee Cloud finding&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I wired Lethe to run on a real Cognee Cloud tenant over its REST API, mapping remember to add plus cognify, recall to search, and forget to the forget endpoint. A live round trip worked: remember a subject, recall their details (leaks), forget, recall again (silent).&lt;/p&gt;

&lt;p&gt;Along the way I hit a real finding that is the whole point of Lethe. Deleting a dataset alone, with DELETE on the dataset, removed the raw data but left the cognified graph still answering with the subject's details. Proper erasure needs the forget endpoint while the dataset still exists. Lethe's Auditor makes that difference visible instead of assuming a delete worked. I reported it upstream as a small pipeline fix as well.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Making it look like a product&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The demo is four core screens plus a compliance cockpit, and it leans into the story rather than hiding the plumbing.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;A live force directed Memory graph, drawn on plain canvas with no libraries, shows customers, cities, and complaint types with their cross links. It can also fetch and render the actual cognified graph from Cognee Cloud, and erased subjects collapse into redacted red ghost nodes.&lt;/li&gt;
&lt;li&gt;A Prove on Cognee Cloud panel runs the real round trip live and shows the before and after answers with PII highlighted.&lt;/li&gt;
&lt;li&gt;A compliance cockpit gives a data protection officer view of every subject's contamination and erasure status, plus an append only ledger of signed certificates that exports to JSON or CSV for a regulator.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;What I learned&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Cognee's lifecycle abstraction did exactly what a good primitive should. It made the memory layer boring, so all my time went to the interesting problem, verifying memory instead of managing it. One dataset per person turned prove this specific human is gone into a simple operation.&lt;/li&gt;
&lt;li&gt;The most convincing demos are honest. The leftover reference finding is more convincing than a fake zero, because it shows the tool is actually looking.&lt;/li&gt;
&lt;li&gt;Inversion is a cheat code for ideas. Everyone is building memory. Building the proof of forgetting was wide open.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Lethe is the river of forgetting in Greek myth. This is the version with receipts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; &lt;a href="https://lethe-two.vercel.app" rel="noopener noreferrer"&gt;https://lethe-two.vercel.app&lt;/a&gt; &lt;br&gt;
&lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/yerramsettysuchita/lethe" rel="noopener noreferrer"&gt;https://github.com/yerramsettysuchita/lethe&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Built on Cognee for the WeMakeDevs and Cognee hackathon. AI coding assistants were used during development. The architecture and the adversarial audit idea are mine.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>hackathon</category>
      <category>vibecoding</category>
    </item>
    <item>
      <title>I Built a Fintech Product in 48 Hours, Got Rejected, and Finished It Anyway</title>
      <dc:creator>Yerramsetty Sai Venkata Suchita</dc:creator>
      <pubDate>Fri, 29 May 2026 19:12:06 +0000</pubDate>
      <link>https://dev.to/yerramsettysuchita/i-built-a-fintech-product-in-48-hours-got-rejected-and-finished-it-anyway-5a86</link>
      <guid>https://dev.to/yerramsettysuchita/i-built-a-fintech-product-in-48-hours-got-rejected-and-finished-it-anyway-5a86</guid>
      <description>&lt;p&gt;&lt;strong&gt;This is a submission for the [GitHub Finish-Up-A-Thon Challenge]&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I Built&lt;/strong&gt;&lt;br&gt;
Ten million property transactions happen in India every year. Every single one of them that involves a loan requires a physical broker to visit the site, write a report by hand, and deliver it to the lender. That process takes three days minimum. Sometimes a week. The loan waits. The business waits. The owner waits. Nobody has questioned this workflow in decades because nobody built an alternative.&lt;/p&gt;

&lt;p&gt;I built the alternative.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Collatiq&lt;/strong&gt; is an AI-powered collateral intelligence engine for Indian NBFCs and lenders. You enter a property address. In 30 seconds you get a lender-grade valuation report with fraud signals, confidence scoring, sensitivity analysis, and a one-click PDF export. No broker. No site visit. No waiting. Zero paid APIs. The entire system runs on OpenStreetMap and Nominatim.&lt;/p&gt;

&lt;p&gt;I submitted it to TenzorX 2026, a national AI hackathon by Poonawalla Fincorp with thousands of competing teams. I did not make the finals. Most people would close the laptop and move on. I opened a new branch instead.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Demo&lt;/strong&gt;&lt;br&gt;
🔗 &lt;strong&gt;Live App:&lt;/strong&gt; &lt;a href="https://collatiq.vercel.app" rel="noopener noreferrer"&gt;collatiq.vercel.app&lt;/a&gt;&lt;br&gt;
💻 &lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/yerramsettysuchita/Collatiq" rel="noopener noreferrer"&gt;github.com/yerramsettysuchita/Collatiq&lt;/a&gt;&lt;br&gt;
🎥 &lt;strong&gt;Demo Video:&lt;/strong&gt; &lt;a href="https://youtu.be/3BJnMP33e6M" rel="noopener noreferrer"&gt;youtu.be/3BJnMP33e6M&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Here is what Collatiq does the moment you submit an address:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;The &lt;strong&gt;Geo Engine&lt;/strong&gt; fires first. It geocodes the address via Nominatim, matches it against one of 34 Bengaluru micromarket zones, and scores infrastructure proximity within a 2km radius using live Overpass API queries against schools, hospitals, metro stations, and major roads.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;The &lt;strong&gt;Valuation Engine&lt;/strong&gt; then applies 12 adjustment factors on top of live government circle rates. Floor premium, age depreciation, construction quality index, road width multiplier, amenity score, FSI utilization, corner plot bonus, and five more. Every factor is derived from real lending practice, not guesswork.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;The &lt;strong&gt;Confidence Engine&lt;/strong&gt; measures how much the system trusts its own output. It scores data completeness, detects anomalies by cross-referencing reported built-up area against zone averages, and surfaces fraud signals with severity classification ranging from critical to informational with human-readable explanations for each flag.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;The &lt;strong&gt;Decision Engine&lt;/strong&gt; produces the final LTV recommendation with sensitivity analysis sliders that let a credit officer model what happens to the valuation if the road width changes, if the property age is different, or if the amenity index shifts. Every slider reruns the full engine in real time with no API call and no page reload.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The output is a structured PDF report that looks like something a credit officer would actually read and file.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Comeback Story&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Here is the honest before and after.&lt;/p&gt;

&lt;p&gt;When I submitted to TenzorX, the four engines existed and the core pipeline ran. But the fraud detection was a placeholder. The sensitivity sliders did not exist. The PDF export printed raw text to the screen. The micromarket data covered 8 zones out of 34. There was no onboarding flow so a first-time user had no idea what to enter or what any of the output meant. The 3D landing page was visually impressive and completely disconnected from the actual application underneath it.&lt;/p&gt;

&lt;p&gt;It was a proof of concept wearing a product's clothes.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What existed at submission&lt;/th&gt;
&lt;th&gt;What was missing&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Four-engine pipeline&lt;/td&gt;
&lt;td&gt;Explainability on every output&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Basic map view&lt;/td&gt;
&lt;td&gt;Interactive zone visualization&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Raw valuation output&lt;/td&gt;
&lt;td&gt;Downloadable lender-grade PDF&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;8 micromarket zones&lt;/td&gt;
&lt;td&gt;34 verified zones with real data&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;No onboarding&lt;/td&gt;
&lt;td&gt;Guided first-use experience&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Placeholder fraud flags&lt;/td&gt;
&lt;td&gt;Real signal-based anomaly engine&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;No sensitivity controls&lt;/td&gt;
&lt;td&gt;Live recalculation sliders&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;After the rejection I rebuilt every one of those gaps.&lt;/p&gt;

&lt;p&gt;The fraud detection engine now cross-references six signal types. Price-to-infrastructure mismatches. Reported area deviations from zone averages. Age versus condition contradictions. Unreported encumbrances inferred from OSM data. Each signal is weighted, classified by severity, and explained in plain language that a loan officer without a technical background can understand and act on.&lt;/p&gt;

&lt;p&gt;The sensitivity sliders were the most technically interesting problem. The naive approach reruns the entire valuation pipeline on every slider change which introduces visible lag. The correct approach memorizes the base valuation, isolates each adjustment factor as an independent multiplier, and recomputes only the delta for the changed factor. The sliders now feel instant because they are instant. The full recalculation takes under 4 milliseconds on a mid-range device.&lt;/p&gt;

&lt;p&gt;The PDF export generates a structured multi-section report using jsPDF with dynamic content layout that handles variable-length fraud flag lists, adjusts section spacing automatically, and produces output that is indistinguishable from a professionally typeset document.&lt;/p&gt;

&lt;p&gt;The micromarket data expansion from 8 to 34 zones required sourcing, verifying, and structuring government circle rate data, infrastructure density measurements, and historical transaction signals for every zone. That is not glamorous work. It is the work that makes the difference between a demo and a product.&lt;/p&gt;

&lt;p&gt;Here is what the numbers look like after the finish-up:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Hackathon submission&lt;/th&gt;
&lt;th&gt;Finished product&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Micromarket zones&lt;/td&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;34&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Valuation adjustment factors&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;12&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fraud signal types&lt;/td&gt;
&lt;td&gt;1 placeholder&lt;/td&gt;
&lt;td&gt;6 real signals&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Report export&lt;/td&gt;
&lt;td&gt;None&lt;/td&gt;
&lt;td&gt;Full structured PDF&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sensitivity controls&lt;/td&gt;
&lt;td&gt;None&lt;/td&gt;
&lt;td&gt;5 live real-time sliders&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;First-use onboarding&lt;/td&gt;
&lt;td&gt;None&lt;/td&gt;
&lt;td&gt;Fully guided flow&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;My Experience with GitHub Copilot&lt;/strong&gt;&lt;br&gt;
I want to be specific about this because most write-ups about AI tools are vague in a way that helps nobody.&lt;/p&gt;

&lt;p&gt;The fraud detection rewrite was the hardest part of the finish-up. The original placeholder was four lines. The real engine needed to reason across six signal types with weighted severity classification and produce explanations that are accurate enough for a credit officer to act on. I knew the architecture I wanted. I wrote a detailed comment block describing the input schema, the signal types, the severity weights, and the expected output format.&lt;/p&gt;

&lt;p&gt;Copilot read that comment and suggested separating signal detection from signal classification into two distinct layers. I had been about to collapse them into a single function. The suggestion was architecturally correct. The separation made the fraud engine testable in isolation, extensible without touching existing logic, and dramatically easier to reason about when the output looked wrong. That one suggestion saved me an entire debugging session that I would have had if I had built it my original way.&lt;/p&gt;

&lt;p&gt;For the jsPDF layout I described the target document structure in a comment and Copilot drafted the coordinate arithmetic for the multi-section layout. PDF generation with dynamic content lengths requires precise y-coordinate tracking across sections and Copilot handled that mechanical complexity while I focused on the content logic and the output formatting.&lt;/p&gt;

&lt;p&gt;For the micromarket data structure I gave Copilot the schema for three zones and asked it to infer the pattern and generate a consistent template for the remaining 31. It produced a starter template that I could verify against government sources and populate. A task that would have taken four hours of repetitive data structuring took 40 minutes of verification and correction.&lt;/p&gt;

&lt;p&gt;The most important thing Copilot did was never let me write boilerplate when I should have been thinking. Every time I started typing something mechanical it finished it. Every time I wrote a comment describing intent it translated intent into structure. That left my cognitive budget entirely available for the decisions that actually mattered.&lt;/p&gt;




&lt;p&gt;Collatiq got rejected from TenzorX. Looking at what it is now versus what I submitted under the deadline I am genuinely glad it happened. The rejection forced me to separate what I had shipped from what I had actually built. They were not the same thing.&lt;/p&gt;

&lt;p&gt;Hackathons teach you to ship under pressure. Finish-up-a-thons teach you what shipping under pressure costs you. This challenge gave me the reason to pay that debt back.&lt;/p&gt;

&lt;p&gt;The product is live. The code is open. The report takes 30 seconds.&lt;/p&gt;

&lt;p&gt;Go try it: &lt;a href="https://collatiq.vercel.app" rel="noopener noreferrer"&gt;collatiq.vercel.app&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Yerramsetty Sai Venkata Suchita&lt;/em&gt;&lt;br&gt;
&lt;em&gt;R&amp;amp;D Intern at Siemens · Indian Patent Filed · IEEE Paper Under Review&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;em&gt;GitHub: &lt;a href="https://github.com/yerramsettysuchita" rel="noopener noreferrer"&gt;@yerramsettysuchita&lt;/a&gt;&lt;/em&gt;
&lt;/h2&gt;

</description>
      <category>devchallenge</category>
      <category>githubchallenge</category>
      <category>githubcopilot</category>
    </item>
  </channel>
</rss>
