Open to backend & full-stack roles
I build the parts of a product that have to be right — warnings that quote the caller's own words, a tripwire that fails closed, tests frozen before the first result. Python on the server, AWS around it, and AI only where code can check its work.
Title blockRev. A
who is drawing this
Server-side by preference, full-stack by necessity.
I'm a computer engineering student who wandered onto the server side and never left. The work I like has a correct answer buried in it somewhere — a warning that has to quote what the caller actually said, a tripwire that fires on the third complaint and not the second, a chain of citations whose risk has to add up.
Most of what I ship is Python, event-driven on AWS more often than not, with a small, sharp frontend on top. I'd rather spend the afternoon on the evaluation than the hover state. (This page is what happens when I get to do both.)
2 hackathon prizes, both solo, and 2 fixes merged upstream — 2026
WeMakeDevs × AWS · Bharat Builds Tour, event 02 · 8–11 Oct 2026
Runner-up
9,146 registered · one of 4 runners-up across all three tracks
Teesri Shikayat — a tripwire for dirty tap water. Hindi voice complaints on Telegram; when three homes close together report it, every home around them is warned, and only the residents can close the case.
WeMakeDevs × Modiqo · 1–6 Sep 2026
3rd place
of ~900 published plays
Four plays + one platform defect — four read-only GitHub plays for Modiqo's Rote registry, and the finding that carried them: the runtime silently kept only the first 64 KiB of a step's output and still reported success. 221 of the 249 public plays that read that output (89%) never checked. Filed upstream; fixed in rote v0.83.0.
numpy/numpyMerged
BUG: close the file in np.load when an .npz is invalid
np.load left the file open when handed a broken .npz — an interrupted download, say — and Python warned about an unclosed file. The file is now handed over only once the archive has opened, so a failure closes it. With a regression test.
python-pillow/PillowMerged
Improve mmap buffer length check
Opening a truncated image by filename crashed with a bare ValueError and ignored LOAD_TRUNCATED_IMAGES: the length check used the tile's stride, which is 0 for PPM, TIFF, TGA and five other formats. It now falls back to the decoder, so a filename behaves exactly like a file object.
| When | Hackathon | What I built | Outcome |
|---|---|---|---|
| Nov 2026 | Flip the SwitchWeMakeDevs × DevTools Academy × Switch | Jaisa Kaha — an AI scribe that won't invent a cook's quantities | Submitted · Track 02 |
| Nov 2026 | Global AI Developer HackathonMeta | Teesra Kaan — warns on a scam call and shows the caller's next line | Submitted · Indie track |
| Oct 2026 | Graph HacksWeMakeDevs × FalkorDB | Shaky Ground — finds the science built on retracted research | Submitted · Track 01 |
| Oct 2026 | Dark FactoryWeAreDevelopers × Band | The Adversary Line — five coding-agent seats building and attacking a reservation service | Submitted |
| Sep 2026 | IBM Bob 2.0 Hackathonlablab.ai × IBM | Antibody — checks that fixed bugs stay fixed (team of 2) | Submitted |
| Sep 2026 | First CommitWeMakeDevs × AWS | Pukaar — one press pages the three people most likely to answer, at once | Submitted · ~1,180 entries |
| Sep 2026 | AI Trading Agents Hackathonlablab.ai × Alpaca | Contour — an options agent that sells only the rich side of the skew | Submitted |
the tools, drawn where they sit in a request
Every tool I reach for, placed at the hop where it does its job — from the browser, through the edge and the application, out to data, background work and models.
Enough frontend to ship the whole thing solo.
Off the laptop, and staying up.
Where most of the time goes.
Storage, cache, and the queries between.
Work that shouldn't block a request.
Used where it earns its latency.
Tests on every push; main deploys on green.
4 builds, opened up
04.1 — Featured build · Runner-up, AWS Environmental Hacks
“the third complaint”
A tripwire for dirty tap water, from my own monsoon tap in Mumbai. Residents complain with a Hindi voice note on Telegram. When three different homes within 250 m report dirty water inside 72 hours, every enrolled home in the ring is warned in Hindi — and only the residents can close the case.
The ward office said resolved. The AI believed them. Cedar didn’t.
The model chooses the words; code and policy choose the actions. A ward office’s “resolved” only asks to close the case — any resident’s “not clean” reopens it, and silence never closes it.
Hindi voice notes arrive on Telegram and go through Amazon Transcribe (hi-IN). Bedrock quota was zero, so an open model — Gemma 4 E4B in llama.cpp on our own EC2 Graviton4 instance — fills a fixed schema, and code validates every field.
DynamoDB Streams feed an EventBridge Pipe into a rule written in plain code, not AI: three different homes, every pair within 250 m, within 72 hours. One DynamoDB transaction settles two “third” reports arriving at once.
One Step Functions execution per incident, waiting on three task tokens: the volunteer’s approval, the ward office’s reply and the residents’ check-ins. A Map state sends every warning as text and as an Amazon Polly voice note.
A Strands case agent whose every action is checked by Cedar policies that fail closed. The ward office gets one email with counts, never names or numbers. Own KMS key, CloudTrail, GuardDuty and alarms.
04.2 — Scam-call guardian · Meta Global AI Developer Hackathon
The third ear on a scam call. Put a suspicious call on speaker and it listens in Hindi, Marathi or English. The moment the caller starts a known scam it warns you in your language, quoting his own words — then shows his next line before he says it.
Ten scam scripts from police, RBI and DoT advisories, written as ordered steps. Once two steps are heard it shows the next one, but only when every script that fits the call agrees. Fixed scripts make the prediction, not a model.
Muse Spark answers 13 fixed yes/no questions about the caller, quoting his exact words for every yes, and a quote that isn't in the transcript is thrown away. A warning needs a quoted sign; red waits for an ask for money, an OTP, an app or a link.
The test was fixed before any result existed and run once: 150 real FTC robocalls and Hindi, Marathi and English calls in synthetic voices, against a keyword list and against the model asked plainly. Every transcript is committed; the run replays offline.
Ten people, five of them elderly, used it in their own language. One elder had lost ₹20,000 to a “sent too much by mistake” call; it became the tenth script before the test was frozen.
04.3 — Citation graph · Graph Hacks 2026, Track 01
Finds the science built on retracted research, not just the papers that cite it. Every citation is weighted by how much the citing paper relies on the cited one, a chain's risk is computed in Cypher, and an agent decides which branches to explore — returning every answer with its evidence.
Another paper citing Paper A:
Paper Arisk 1.0
Background× 0.15mentions it in passing
Paper Drisk 0.15In passing
Builds on ⇔ risk ≥ 0.2: three supporting links score 0.6³ = 0.216, and any background link caps a chain at 0.15.
The model never picks a number. It reads only the citation sentences and answers yes/no questions; fixed rules turn the answers into a class and a weight, and the deciding sentence is stored verbatim on the edge.
Risk flows back from the retracted paper one hop per query, so the cost grows with edges, not chains. A variable-length path with reduce() returns one paper's riskiest chain, evidence on every link, in about 20 ms.
A second link counts only if it relies on the part that came from the retracted paper. One model wasn't reliable enough, so two model families vote: a link carries only if both say yes.
A web UI, a foundation check for any paper or a whole reference list, and an MCP server so an AI agent can check a paper before citing it. Unchecked is never shown as clean.
04.4 — Agents with owners · Flip the Switch, Track 02
An AI scribe that refuses to make up a cook's quantities. It writes down only the amounts that were said and keeps a gap where none was — or where the cook said “by feel” — and only the cook's own account can fill it. On a test sentence, four models out of four turned “ghee, by feel” into a number.
shipped work, newest first
Final-year computer engineering student, so no employment history yet — what follows is shipped, dated work.
Nov 2026
Flip the Switch · WeMakeDevs × DevTools Academy × Switch · solo
An AI scribe for a family's recipes that writes down only the amounts the cook said and keeps a gap where none was. Three agents share one Discord room through Switch, and the cook's own agent refuses anyone else who tries to fill a gap — the AI, and me.
Oct–Nov 2026
Meta Global AI Developer Hackathon · Indie track
A scam-call guardian in Hindi, Marathi and English that warns in the caller's own words and shows his next line before he says it. Tested once, on a set fixed before any result: 131 of 132 real FTC impersonation robocalls warned on.
Oct 2026
Graph Hacks · WeMakeDevs × FalkorDB · solo
Finds the science built on retracted research: a FalkorDB citation graph where every citation is weighted by reliance and a chain's risk is computed in Cypher, with an agent that decides where to dig and an MCP server for other agents.
Oct 2026
WeMakeDevs × AWS Environmental Hacks · soloRunner-up
A dirty-water tripwire for Mumbai: Hindi voice complaints on Telegram, a plain-code rule for three homes within 250 m inside 72 hours, and one Step Functions case per incident that only residents can close, guarded by a Strands agent checked with Cedar. One real incident cost ₹0.94.
Oct 2026
Dark Factory · WeAreDevelopers × Band · solo
A dark software factory: five coding-agent seats in Band Desktop building a reservation service in stages, each a complete container, while one seat plants known bugs to test the checker. Run on free models, for $0.
Sep 2026
IBM Bob 2.0 Hackathon · lablab.ai · team of 2
Do fixed bugs stay fixed? Antibody puts a project's past bug fixes back, one at a time, and runs its own tests. On sqlparse, 13 of 20 fixes were caught before and 18 of 20 after. Before the event: 479 fixes undone across 18 Python libraries, and 139 times every test still passed.
Sep 2026
WeMakeDevs × AWS First Commit · solo
Help for an older person living alone: one press pages the three people most likely to answer, at the same time; the first to say “I'm going” wins, and the circle widens if nobody answers within a minute. Eleven languages, AAA contrast, 21 live checks.
Sep 2026
WeMakeDevs × Modiqo Rote Playoffs · solo3rd place
Four read-only plays for pull-request review — unresolved threads, flaky CI, who should review, and a truncation audit — plus a measured platform defect: step output silently cut at 64 KiB, unchecked by 89% of public plays. Filed upstream and fixed in rote v0.83.0.
Aug–Sep 2026
AI Trading Agents Hackathon · lablab.ai × Alpaca
An autonomous options agent that measures the volatility surface every 15 minutes and sells only the rich side of the skew, on paper. A 387-cycle backtest found no edge (t = +0.37), so none is claimed. 302 tests, and every decision recorded in a SHA-256 hash chain.
I'm after backend or full-stack work where the hard part is the system design, the data model, or keeping a deploy boring. If that's what you've got, I'd like to hear about it. Fastest route is email — I reply to everything that isn't a recruiter template.
Local time, Mumbai, India IST · UTC+5:30
