Rishabh Bansal

Fremont, California / Incoming UIUC

Rishabh Bansal.

I build software for when the easy path fails: no cell service, twelve hours of footage, a robot in a new room, or a bug nobody can reproduce.

See selected work hello@rishabhb.dev

01 / Selected work

Six builds worth opening.

The short version is impact. The build notes hold the architecture, my contribution, and the part that is still honestly unfinished.

01 / Flagship / Disaster response

AnchorMesh SafeRelay

Move an SOS packet through nearby phones when cell towers are unavailable, then give responders a source-backed operations view when any relay gets online.

1st place
Alameda Hacks
1st place
JacHacks Social Impact
Offline first
BLE store + forward

What I built

I led the product and engineering work across the rescue console and the SafeRelay rebuild: Jac cloud ingestion, the CoreBluetooth bridge, packet validation and deduplication, background sync, local notifications, and the evidence model behind uploads, acknowledgements, and outcomes.

Build notes

The mobile client stores packets locally, advertises and scans over BLE, limits hops and expiry, and uploads idempotently when connectivity returns. The web side combines verified relay evidence with public USGS, NWS, NOAA, and GDACS context.

I built it with Jerry Wen and Aditya Das. The original won Alameda Hacks in 2025; the Jac, Capacitor, and native iOS rebuild won the Social Impact track at JacHacks in 2026.

Evidence boundary: a browser demo or protocol test does not prove radio delivery. End-to-end multi-hop remains gated on captured three-phone field evidence.

Open SafeRelay source
AnchorMesh demo showing an SOS packet moving through phone, relay, volunteer, gateway, and command nodes
A clearly labeled simulation of the relay path. It never substitutes demo state for field evidence.

02 / Animal welfare / Video intelligence

ZooVision

Turn overnight camera footage into reviewable moments, deterministic triage, and a keeper handoff instead of another wall of video thumbnails.

3rd place
AWS Builder Loft
10 sources
One review workspace
0 model votes
On severity

What I built

I led the end-to-end system: TwelveLabs upload analysis, strict provider schemas, wall-clock timestamp normalization, cross-chunk stitching, daytime per-animal baselines, deterministic Python triage, SQLite and Neo4j persistence, and a grounded assistant whose answers jump to the right camera and moment.

Build notes

Models can describe what appears in a scene. They cannot assign severity, diagnose an animal, or recommend treatment. First-match Python rules own alert levels; failed or invalid provider calls become visible DataGaps.

The keeper workspace connects playable footage, provider observations, event timelines, a live Neo4j graph, acknowledgements, outcomes, and a morning brief. Jerry Wen and Aditya Das built it with me.

Evidence boundary: ZooVision supports review. It is not a medical device and does not claim behavior-recognition accuracy without labeled evaluation footage.

Visit ZooVision Source
ZooVision animal welfare monitor with camera feeds, event evidence, and alert workflow
The live review surface connects footage, timestamps, policy, and keeper action. Nox keeps watch.

03 / Robotics / Measured learning

Muscle Memory

Keep one robot fixed while its worlds, experience, and high-level task policy evolve. Every candidate has to earn promotion on unseen homes.

2nd place
Best Use of Guild.ai
37 / 40
Commits authored
4.5 hours
Live build window

What I built

I led the system integration and product: validated seeded apartments, immutable episode records, FalkorDB experience memory, Guild.ai specialist review, RocketRide execution gates, the live household simulator, and the deployment path that keeps policy claims tied to evaluation receipts. Git history attributes 37 of the repository's 40 commits to me.

Build notes

MM-01 carries a medicine pouch through changing MuJoCo apartments. A frozen controller handles balance and walking; the learned policy may only choose speed, turn rate, and stop probability. LaserData stores 20 Hz telemetry and FalkorDB connects failures, lessons, worlds, and policy versions.

Promotion requires at least 80% success across 20 isolated worlds, zero falls, bounded collisions, safe clearance, and a material lift over baseline. I built it with Aditya Das.

Evidence boundary: the stronger candidate improved development success from 25% to 66.7%, but two falls kept it out of held-out evaluation. No learned checkpoint is presented as promoted yet.

Try the live experience Source
Muscle Memory household simulator with MM-01, a delivery task, sensor readouts, and policy evidence
One body, many worlds, and a promotion gate that keeps failed policies visible.

04 / Developer tools / Local first

QAgent

Reproduce a web bug, repair it in an isolated worktree, rerun the browser check, and leave enough evidence for a human to trust the change.

Winner
Best Use of Browserbase
3 surfaces
Desktop, CLI, MCP
1 engine
Durable local state

What I built

I designed the Tester → Triage → Fixer → Verifier loop, then rebuilt it as one local-first engine powering Electron, CLI, and MCP. Runs survive restarts in SQLite; patches happen in isolated Git worktrees; screenshots, DOM evidence, logs, and checksums stay attached to the result.

Build notes

QAgent detects project commands, opens a real Chromium session, captures a failing flow, asks a configured model to diagnose and patch it, verifies the repair in a fresh run, and can publish only through explicit GitHub policy.

Dirty checkouts and sensitive changes block publication. Browser, model, GitHub, and tracing providers fail visibly instead of silently swapping in fake output. The current release is public under AGPL-3.0.

Open QAgent
QAgent desktop app showing a verified repair and checksummed browser artifacts
A real repair run, with the before/after browser evidence still attached.

05 / Assistive technology / Research

SmartCane

A roughly $70 cane prototype that combines obstacle sensing, fall detection, thermoelectric safety lights, and a Bluetooth caregiver app.

97.75%
Distance accuracy in bench trials
27 / 27
Simulated falls detected
1st place
County science fair + Stem4All

What I built

I co-designed the Arduino hardware and sensor stack, built the Bluetooth Android caregiver flow, ran the measurement study, and coauthored the COSITE 2025 paper with Jasamarbir Arora and Krishna Durgavarjhula. The work also received an accepted provisional utility patent.

Build notes

Ultrasonic sensors measure nearby terrain and obstacles; a gyroscope triggers a fall workflow; a pulse sensor and HC-05 feed the mobile app; temperature differentials power LEDs through thermoelectric modules.

The paper reports 97.75% ± 1.2% obstacle-distance accuracy up to 1.5 meters, all 27 simulated falls detected, no false positives in 18 non-fall trials, and heart-rate MAE of 3.10 BPM against a consumer oximeter.

Evidence boundary: these were preliminary bench and simulated trials, not a clinical study or evaluation with elderly users.

Page from the SmartCane COSITE paper showing the prototype, component comparison, and terrain tests
COSITE 2025: the prototype and its lab evidence, shown together.

06 / Generative media / Cloud systems

Kinora

Turn a chapter into a sequence of accepted film shots, while keeping character, setting, and story continuity visible to the reader.

Qwen + Wan
Planning and video
Shot gates
Measured before acceptance
Alibaba Cloud
End-to-end deployment

What I built

I led the reader, generation pipeline, cloud deployment, and quality gates. Kinora plans scenes with Qwen, renders clips with Wan, normalizes them for browser playback, and preserves the accepted live shot rather than quietly replacing a failed render with a local placeholder.

Build notes

The system separates story planning, reference-conditioned generation, measured visual checks, acceptance, and playback. Deployment tooling runs the API, workers, queue, storage, and reader on Alibaba Cloud, with explicit reset and health paths.

I built it with Jerry Wen for the Qwen hackathon. The project was submitted and deployed; no award outcome is claimed here.

Open Kinora
Kinora reader showing Alice in Wonderland with a generated cinematic scene
A book stays readable while its accepted scenes become film alongside it.

02 / Experience

Work that lasted longer than a weekend.

Shipping, maintenance, distribution, and the unglamorous details that make software usable.

July 2024 – September 2025

Software Engineer Intern

Hey, Blue!

Work supported a platform reporting 500+ police-community connections across 14 states.

Product + data

Moved connection data into a Supabase-backed live map, added expandable announcements, and improved app links and QR onboarding so people could reach the right mobile flow.

Trust + safety

Implemented mandatory phone verification for purchases, modeled SMS operating cost, added clear unverified states, and helped constrain identity-sensitive profile photos to a live front-camera path.

iOS delivery

Fixed a mobile crash, configured permissions and associated domains, handled signing and distribution work, and wrote the team's practical Xcode, CocoaPods, EAS, TestFlight, and App Store release runbook.

September – December 2023

AI Creator

Lunar Coffee Labs

Produced weekly AI-assisted brand content and iterated prompts against creative feedback. It was my first sustained role translating new model capabilities into repeatable output for a real team.

Earlier labs, clearly labeled

2025 / UC Berkeley Summer Springboard

Data science program

Built and evaluated an RNN-based spam classifier.

2024 / Inspirit AI

AI program

Worked on an exoplanet-detection project with scientific data.

2023–24 / MIT Beaver Works

Assistive technology program

Developed the early problem framing that became SmartCane.

2025 / Bay Area Cyber League

Cybersecurity program

Practiced team-based defense and security problem solving.

03 / More work

The rest of the build log.

Smaller descriptions, same standard: what it did, what I contributed, and what happened.

2026 / 1st place

FairValue

Co-built a multiplayer real-estate prediction market with Sean Chiu: live rooms, QR joining, Over/Under appraisals, and an LMSR market maker that kept prices coherent. Won the Cognee AI-Memory Hackathon.

Source

2026 / Best Use of InsForge

MasterBuild

Built a six-agent research and planning system that turns a raw product idea into a source-backed build plan using shared memory, realtime state, and parallel web research.

2026 / Local first

Mnemo

Built a native Swift macOS memory workspace with a notch interface, local Ollama models, an egress audit, and nine ownership-partitioned agents. Submitted to Supermemory Localhost; no award result is claimed.

2026 / Finalist

FlowPR

Shaped QAgent's repair loop into a hosted frontend agent with live run state, browser artifacts, health gates, patch verification, and proof-backed pull requests.

Source

2026 / Human in the loop

ImmiFile

Built a browser agent for immigration filing with PII redaction, review gates, portal navigation, document extraction, and reusable profile memory for repeated forms.

Source

2026 / Incident response

RePro

Built a verification-first remediation agent that ingests production evidence, reproduces a failure in a sandbox, patches it, runs gates, and opens a receipt-backed PR.

Source

2026 / Privacy utilities

FreeYT + Navio

Shipped two local-first Safari extensions: one routes YouTube through no-cookie embeds; the other opens Google Maps links in Apple Maps without collecting user data.

Source

Always / Tiny tools

Side quests

Browser experiments, market-data tools, native utilities, deployment scripts, and the occasional project that exists because a small annoyance lasted one day too long.

GitHub

04 / About

I care about the part after the demo.

I am an incoming University of Illinois Urbana-Champaign student studying Information Sciences + Data Science. I like projects where software has to meet a physical or human constraint: radios go to sleep, model output is uncertain, robots fall, and users do not owe the interface patience.

That is why my projects keep returning to local-first state, deterministic policy, receipts, visible failure, and real operator workflows. I move fast at hackathons, but I do not think speed is an excuse to hide what is simulated, unverified, or still broken.

I have been playing guitar for years, enjoy a slightly absurd side project, and do my best work with people who care about both the system and the person stuck using it.

05 / Contact

Have a hard problem with a real edge case?

hello@rishabhb.dev
GitHub LinkedIn