Ayush Madhav

Jaseci Labs

Software Engineering Intern

• Engineering a domain-specialized coding agent for the Jac language, fine-tuning Gemma 4 26B with Unsloth and 4-bit QLoRA on an object-spatial superset of Python where general-purpose models have no usable priors.
• Architecting a multi-recipe synthetic-data pipeline (Python to Jac translation, adversarial DPO negatives, evol-instruct, self-distillation) targeting 300k+ examples behind a compiler and unit-test hard-gate.
• Designing a four-stage training curriculum across core SFT, debugging and reasoning, DPO alignment, and multi-turn agentic data to lock in idiomatic walker, node, and edge constructs and prevent Python-pattern fallback.

University of Michigan, College of Engineering

Student Researcher (POSE)NSF, $300K

• PhD-level research as an undergrad on NSF Phase I POSE, growing an open-source ecosystem around the data-spatial Jac language.
• Building the project's digital infrastructure, docs site, CI, and contributor onboarding tooling to lower the barrier to entry and grow a distributed contributor community.
• Authoring tutorials and educational materials, and analyzing developer feedback and usage telemetry to refine onboarding and inform long-term governance models.

Cactus

Core ContributorYC S25

• Core contributor on an open-source mobile AI inference engine for smartphones and low-power devices.
• Implementing ARM SIMD kernels for matmul and attention with KV-cache quantization, and zero-copy computation graphs for on-device transformer inference at sub-50ms time-to-first-token.
• Integrating NPU acceleration and INT4 quantization, plus streaming inference and cloud-handoff logic, and building OpenAI-compatible FFI APIs for Flutter, React Native, and Kotlin.

CareTether

Co-Founder, Frontend and OutreachPatch accelerator

• Co-founded a wearable and app system for seniors with cognitive difficulties, built on-site at the Patch entrepreneurship accelerator.
• Piloted with 5+ families and reached 20,000+ people on social media, and pitched venture possibilities to NGOs and care homes.

GIIT Solutions

Associate Software Intern

• Maintained CI/CD pipelines on AWS, automating deployments across 12+ microservices and improving reliability by 28%.
• Developed Python scripts to streamline backend processes, reducing manual overhead by 40% and saving 15+ hours weekly.
• Optimized cloud infrastructure with senior devs, achieving a 35% cost reduction through resource management.
• Deployed cloud-native solutions with AWS CI/CD tools, improving scalability and achieving 99.7% uptime.

MK-OSINT

A real-time geospatial intelligence platform that plots live public feeds (military aircraft transponders, the ISS, USGS earthquakes, Safecast radiation sensors, airport control zones) on a CesiumJS 3D globe with toggleable layers, entity history, and CRT/night-vision/FLIR filters. A TypeScript backend runs a declarative ingestion engine: each source is one YAML file (fetch, parse JSON/GeoJSON/XML/CSV, map fields), polled with retry/backoff, stored in SQLite, and pushed over WebSocket with an OpenAPI-documented REST API. Also ships as a spec-driven AI-agent workshop kit, including an agent skill that turns any public data URL into a source definition.

Jac ML Studio

A full-stack local ML workbench and research pipeline for finetuning LLMs on Jac (Jaseci Labs' object-spatial programming language), pairing a synthetic-data SFT/DPO/GRPO training pipeline (MLX, LoRA, Qwen3-Coder-30B) with a pure-Jac web app — server endpoints and a React-in-Jac client from one codebase — for chatting with trained models, running training jobs, and tracking evals.

NightShift

An unattended nightly automation harness that syncs a Jaseci fork, runs deterministic lint/format passes, then dispatches Claude agent sessions to audit and apply small verified fixes/refactors before opening PRs, orchestrated via bash, launchd, and native Jac scripts.

NASA SUITS Challenge

AI Lead for CLAWS UM's NASA SUITS entry. Autonomous lunar-EVA navigation with A*/Jump Point Search over a 3-layer occupancy grid, 17-ray LiDAR fusion, and RSSI trilateration; PyTorch obstacle detection (temporal ResidualBeamCNN, BEV cross-attention transformer) behind a hot-swappable inference framework; and a voice pipeline streaming 16 kHz audio over WebSocket to faster-whisper and a neural intent classifier.

Red Bull Kappa Theta Pi Hackathon

1st Overall, Education Track

Turns exam results and concept maps into readiness insights. It clusters struggling students, traces root-cause concepts, and gives instructors an interactive prerequisite graph that shows dependencies and gaps at a glance.

MLH RevolutionUC

Winner, ElevenLabs Track

A multimodal medical-bill parser with a dual-path backend. It returns structured line items, savings signals, and financial-aid screening, generates editable PDF appeal packets, and runs a live call assistant over WebSocket.

Anthropic Claude Hackathon

Winner, Audited AI Track

A browser extension that reads privacy policies, cookie notices, and extension permissions, assigns a privacy-risk score with Claude, and surfaces the warnings that matter so users can see how their data is handled.

Jaseci Hackathon

2nd Overall · Cash prize

Jac-powered security and PR review. Verified GitHub webhooks fire graph walkers that map the codebase, score dependency and blast-radius impact, and surface risk in a web control room. Built to advance open-source dev tooling.

24-Hour AI + Business Hackathon

Top 5 Finish

An AI concept-readiness platform for instructors and students. Instructors upload exam scores and a concept graph and get a class heatmap with root-cause traces. Students get a personal report with a prerequisite-ordered study plan.